config

package
v0.6.9 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Jul 13, 2026 License: AGPL-3.0 Imports: 11 Imported by: 0

Documentation

Index

Constants

View Source
const DefaultHomeHeader = "X-Memini-Home"

DefaultHomeHeader is the request header carrying the caller's personal namespace (see Config.Home). Fixed (no env override), same rationale as DefaultNamespaceHeader. Unlike the namespace header, its absence has no default — no header means no home leg for that request.

View Source
const DefaultNamespaceHeader = "X-Memini-Namespace"

DefaultNamespaceHeader is the request header carrying the tenant namespace. Fixed (no env override): clients and plugins all send this exact header.

Variables

This section is empty.

Functions

func DeprecationWarnings added in v0.5.0

func DeprecationWarnings() []string

DeprecationWarnings returns one message per removed, non-fatal environment variable that is currently set, telling the operator what to use instead. These variables are ignored either way; this only explains the change. Empty when none are set. Fatal deprecated vars (see FatalDeprecatedVars) are excluded — they refuse the boot instead of just warning, so a warning here would never be reached in that path anyway.

func FatalDeprecatedVars added in v0.6.6

func FatalDeprecatedVars() []string

FatalDeprecatedVars returns one refusal message per fatal deprecated environment variable (MEMINI_GLOBAL_NAMESPACE, MEMINI_TENANT_SHARED) that is currently set. Unlike DeprecationWarnings, these are not safe to boot through silently: both named the old opt-in shared-scope model, which the always-on ancestor cascade replaced outright, so booting as if the var were never set would silently drop the operator's expectation of shared visibility. Empty when neither is set.

Deliberately NOT checked inside Load(): `memini migrate scopes` (cmd/memini/migrate.go) also calls config.Load() and separately reads MEMINI_GLOBAL_NAMESPACE via os.Getenv to print adoption instructions for exactly this case. If the refusal lived in Load(), the one command that handles the migration could never run while the var that triggers it is set — an operator deadlock. Instead this is called explicitly from the server-boot paths (cmd/memini/root.go runServer and runMCP), the callers that need the refusal: booting a long-running server (REST or MCP) with a stale shared-scope expectation baked into the operator's env is the dangerous case; one-shot CLI commands (migrate, doctor, reembed, ...) are not "booting" anything and are unaffected.

func NamespaceOverride added in v0.6.9

func NamespaceOverride(dir string) (string, bool)

NamespaceOverride returns the per-project namespace override in force for dir, and whether one exists. Any read or parse error yields ("", false) — a broken overrides file must degrade to automatic resolution, never fail a command.

func OverridesPath added in v0.6.9

func OverridesPath() string

OverridesPath exposes the overrides file location so `memini doctor` can point at it when reporting that an override is in force.

Types

type Backend

type Backend string

Backend selects the storage driver.

const (
	BackendSQLite   Backend = "sqlite"
	BackendPostgres Backend = "postgres"
)

type Config

type Config struct {

	// HTTPAddr is the address the main listener binds. It carries the REST API,
	// the MCP endpoint, and the admin UI unless UIAddr moves the UI elsewhere.
	HTTPAddr string `env:"MEMINI_HTTP_ADDR" envDefault:":8080"`
	// ShutdownTimeout is how long in-flight requests get to finish after SIGTERM
	// before the server stops waiting and exits anyway.
	ShutdownTimeout time.Duration `env:"MEMINI_SHUTDOWN_TIMEOUT" envDefault:"15s"`
	// RequestTimeout bounds how long a single /v1 REST request may run before
	// chi's Timeout middleware cancels its context (internal/api/rest.Mount).
	// It never applies to /mcp (long-lived SSE), /healthz, /readyz, or
	// /metrics. Default 60s rather than a more conservative 30s: the LLM HTTP
	// client's own timeout is 120s (internal/llm/llm.go defaultHTTPTimeout),
	// and POST /v1/answer can ride that full call chain (e.g. a
	// reasoning_level=high rewrite/answer). A 30s default would systematically
	// cut off those legitimate long-running answer calls; 60s cuts that risk
	// roughly in half without going as high as the LLM client's own ceiling.
	// It still doesn't fully cover a 120s LLM call — deployments that
	// regularly hit that ceiling should raise this explicitly. 0 disables it.
	RequestTimeout time.Duration `env:"MEMINI_REQUEST_TIMEOUT" envDefault:"60s"`
	// MetricsAddr, when set (e.g. ":9090"), serves /metrics on its own listener
	// instead of the main HTTP port. The dedicated port is meant to stay
	// in-cluster — keep it out of any public route and it needs no bearer token.
	// Empty (the default) keeps /metrics on the main port, where MEMINI_API_KEY
	// gates it.
	MetricsAddr string `env:"MEMINI_METRICS_ADDR"`
	// UIAddr, when set (e.g. ":8081") and distinct from HTTPAddr, serves the
	// admin UI on its own listener instead of the main HTTP port. The shell
	// embeds MEMINI_API_KEY, so isolating it to a dedicated port lets operators
	// expose that port only on a trusted (LAN) gateway while the main port
	// carries the token-free API. The UI listener also serves the API so the
	// same-origin SPA can call /v1. Empty (the default) keeps the UI on the main
	// port when MEMINI_UI_ENABLED is true.
	UIAddr string `env:"MEMINI_UI_ADDR"`

	// LogLevel is one of debug, info, warn or error.
	LogLevel string `env:"MEMINI_LOG_LEVEL" envDefault:"info"`
	// LogFormat is json (structured, for a log pipeline) or text (readable, for
	// a terminal).
	LogFormat string `env:"MEMINI_LOG_FORMAT" envDefault:"json"`

	// Backend selects the storage driver: sqlite (embedded, the default, no
	// external dependency) or postgres (pgvector/VectorChord, for a shared
	// deployment). Postgres additionally requires PostgresDSN.
	Backend Backend `env:"MEMINI_BACKEND" envDefault:"sqlite"`
	// SQLitePath is where the embedded database file lives. Give it a path on a
	// volume that survives restarts; the default is relative to the working
	// directory, which in a container usually means it does not.
	SQLitePath string `env:"MEMINI_SQLITE_PATH" envDefault:"memini.db"`
	// PostgresDSN is the connection string used when Backend is postgres, e.g.
	// postgres://user:pass@host:5432/memini?sslmode=disable. Required in that
	// mode; the server refuses to start without it.
	PostgresDSN string `env:"MEMINI_POSTGRES_DSN"`

	// EmbedBaseURL is an OpenAI-compatible /embeddings endpoint, which you
	// deploy: text-embeddings-inference, llama.cpp, vLLM, OpenAI itself, or
	// anything else speaking that shape. Leaving it empty is supported and
	// degrades recall to keyword-only search, which works but retrieves
	// noticeably worse.
	EmbedBaseURL string `env:"MEMINI_EMBED_BASE_URL"`
	// EmbedAPIKey is the bearer token for the embeddings endpoint, when it wants
	// one. Optional.
	EmbedAPIKey string `env:"MEMINI_EMBED_API_KEY"`
	// EmbedModel is the model name sent with each embeddings request. memini
	// records which model produced a store's vectors and refuses to start when
	// this later disagrees, because vectors from different models are not
	// comparable and a silent swap degrades recall with no error. Use the
	// `memini reembed` command to migrate a store, or ReembedOnModelChange to do
	// it automatically at startup.
	EmbedModel string `env:"MEMINI_EMBED_MODEL" envDefault:"text-embedding-3-small"`
	// EmbedDims is the dimensionality of the embedding model, and it must match
	// the model EmbedBaseURL actually serves. This is the most common setup
	// mistake: the default suits text-embedding-3-small, so pointing at a 768 or
	// 1024 dimension model without changing this corrupts the store rather than
	// failing cleanly. Unlike the model name, dimensionality cannot be migrated
	// in place; changing it needs a fresh store (export, then import).
	EmbedDims int `env:"MEMINI_EMBED_DIMS" envDefault:"1536"`
	// EmbedQueryPrefix is prepended to recall queries before embedding, for
	// instruction-tuned asymmetric embedders (e.g. Qwen3-Embedding, bge).
	// Documents are always embedded without it. Empty disables.
	EmbedQueryPrefix string `env:"MEMINI_EMBED_QUERY_PREFIX"`
	// EmbedMaxBatch caps items per /embeddings request so bulk callers (dedup
	// over a whole namespace) can't exceed the server's max client batch and
	// fail with 422. The TEI default is 32; 20 leaves headroom.
	EmbedMaxBatch int `env:"MEMINI_EMBED_MAX_BATCH" envDefault:"20"`
	// EmbedMaxBatchChars caps total characters per request (0 disables).
	EmbedMaxBatchChars int `env:"MEMINI_EMBED_MAX_BATCH_CHARS" envDefault:"24000"`
	// EmbedMaxConcurrency caps in-flight calls to the embeddings backend. 0
	// is unbounded. Set to 1-2 for self-hosted backends that can't service a
	// recall burst in parallel.
	EmbedMaxConcurrency int `env:"MEMINI_EMBED_MAX_CONCURRENCY" envDefault:"0"`
	// ReembedOnModelChange makes the server re-embed every stored memory at
	// startup when MEMINI_EMBED_MODEL differs from the model the vectors were
	// produced with, instead of refusing to start. Off by default: re-embedding
	// hits the embeddings endpoint once per memory and blocks startup, so it
	// must be opted into (the `memini reembed` command is the explicit
	// alternative). Dimensionality still cannot change this way.
	ReembedOnModelChange bool `env:"MEMINI_REEMBED_ON_MODEL_CHANGE" envDefault:"false"`

	// WriteDedupScore is the fused vector similarity (0..1) at or above which a
	// fresh write is treated as a near-duplicate of its nearest same-tier memory,
	// triggering WriteDedupAction. 0 disables write-time dedup regardless of the
	// action. The right value is embedder-dependent (~0.9 collapses near-identical
	// restatements only; the default 0.625 was calibrated for merge hints in
	// bench/dedup_test.go). See WriteDedupAction for what happens at/above it.
	WriteDedupScore float64 `env:"MEMINI_WRITE_DEDUP_SCORE" envDefault:"0.625"`

	// WriteDedupAction picks what happens when a write scores >= WriteDedupScore
	// against its nearest same-tier memory:
	//   - "hint" (default): store the write and return a MergeHint so the caller
	//     (agent or human) can merge via memory_update. Non-destructive; scoped to
	//     durable tiers (semantic/procedural), where the threshold was calibrated
	//     and the hint is consumed — episodic/working writes skip the lookup.
	//   - "coalesce": reinforce the existing memory and drop the write (headless
	//     corpus hygiene; use a high score like 0.9). Applies to all tiers.
	//   - "supersede": store the write and tombstone the old memory ("new wins").
	//   - "off": no write-time dedup (the exact-restatement fingerprint pass,
	//     WriteDedupFingerprint, still runs independently).
	WriteDedupAction string `env:"MEMINI_WRITE_DEDUP_ACTION" envDefault:"hint"`

	// SplitDedupLLMMerge (opt-in, default off) routes ambiguous split-dedup
	// candidates (≥2 close neighbours) through the LLM consolidator for a
	// merge/supersede verdict before the deterministic action fires. Requires
	// a consolidator to be configured.
	SplitDedupLLMMerge bool `env:"MEMINI_SPLIT_DEDUP_LLM_MERGE" envDefault:"false"`

	// ContradictionDownrank (default on) invalidates a durable fact when a fresh
	// durable write contradicts it (changed value or flipped polarity, confirmed
	// by the lexical detector): the stale fact's valid_to is stamped so it leaves
	// live recall while AsOf time-travel can still reach it, and its confidence
	// is shrunk. Reversible (Restore clears valid_to) and precision-first (0
	// restatement misfires measured in bench/contradiction_test.go). Set false to
	// disable — the kill-switch; there is no threshold knob.
	ContradictionDownrank bool `env:"MEMINI_CONTRADICT_DOWNRANK" envDefault:"true"`

	// Cascade (default true) is the server-wide switch for the ancestor/home/
	// link read cascade. When true, a recall or briefing in namespace N also
	// reads N's ancestors, the caller's home namespace, and N's stored links —
	// durable tiers only. Set false to restore pre-cascade isolation: the
	// default read set becomes N (and its subtree, when asked) only, and Scope
	// "full"/"everywhere" no longer add the cascade legs. A per-call Scope of
	// "project" already suppresses the cascade for one request; this is the
	// global default for operators (or upgraders) who want isolation without
	// setting Scope on every call. See docs/scopes.md#knobs.
	Cascade bool `env:"MEMINI_CASCADE" envDefault:"true"`

	// LLMBaseURL is the master switch for everything that needs a chat model.
	// Empty (the default) is a fully supported way to run: marker heuristics
	// still do write-time extraction, tier classification, promotion,
	// corroboration and contradiction handling, so durable knowledge still
	// accumulates. Setting it adds background consolidation, POST /v1/answer,
	// the memory_answer MCP tool, and Rerank="llm".
	LLMBaseURL string `env:"MEMINI_LLM_BASE_URL"`
	// LLMAPIKey is the bearer token for the LLM endpoint, when it wants one.
	// Optional.
	LLMAPIKey string `env:"MEMINI_LLM_API_KEY"`
	// LLMModel is the chat model used for consolidation, distillation, answering
	// and LLM reranking.
	LLMModel string `env:"MEMINI_LLM_MODEL" envDefault:"gpt-4o-mini"`
	// LLMAPI selects the chat backend: "openai" (default) or "anthropic".
	LLMAPI string `env:"MEMINI_LLM_API" envDefault:"openai"`

	// Rerank selects recall reranking: "off" (default), "llm" (reorder with the
	// chat LLM), or a cross-encoder /rerank base URL (e.g. http://host:8002/v1).
	// Reranking reorders the top k composite-ranked candidates; it adds one
	// reranker call per recall.
	Rerank string `env:"MEMINI_RERANK" envDefault:"off"`
	// RerankModel names the cross-encoder to use when Rerank is a URL. The
	// server warns at boot if a URL is set without one.
	RerankModel string `env:"MEMINI_RERANK_MODEL"`
	// RerankAPIKey is the bearer token for the cross-encoder endpoint, when it
	// wants one. Optional.
	RerankAPIKey string `env:"MEMINI_RERANK_API_KEY"`
	// RerankPool is how many composite-ranked candidates are handed to the
	// reranker before the result is truncated to the recall limit. 0 (the
	// default) reranks exactly the limit, which reorders the results but can
	// never surface a memory that ranked below them — most of a cross-encoder's
	// value is precisely that rescue, so a deployment with a reranker wants this
	// set (RecallPoolSize, ~50, is the natural ceiling: recall never retrieves
	// more). Cost is linear — one model forward pass per candidate — so a deep
	// pool trades recall latency for accuracy.
	RerankPool int `env:"MEMINI_RERANK_POOL" envDefault:"0"`
	// RerankMaxBatchChars caps the total characters across the query and all
	// documents in a single /rerank request. This is an HTTP payload guard, not
	// a context-window guard: a Cohere-style /rerank server scores each
	// (query, document) pair in its own forward pass, so the model's context
	// bounds a single pair, never the batch. Sizing this near the model context
	// shards a deep RerankPool into many *serial* requests (see
	// rerank.CrossEncoder.Rerank), which is far more likely to blow RerankTimeout
	// than a large body is to trouble the server. 0 disables proactive batching.
	RerankMaxBatchChars int `env:"MEMINI_RERANK_MAX_BATCH_CHARS" envDefault:"6000"`
	// RerankTimeout bounds a single reranker call; past it, recall degrades to
	// composite order instead of stalling on a slow or congested backend.
	RerankTimeout time.Duration `env:"MEMINI_RERANK_TIMEOUT" envDefault:"10s"`
	// RerankMaxConcurrency caps in-flight rerank calls. 0 is unbounded. See
	// EmbedMaxConcurrency for the rationale.
	RerankMaxConcurrency int `env:"MEMINI_RERANK_MAX_CONCURRENCY" envDefault:"0"`
	// RecallEmbedTimeout bounds the query embed on the recall path; past it, or on
	// any embed error, recall degrades to keyword-only search instead of stalling
	// on a slow or stuck embeddings backend. Defaults to 2s so a wedged backend
	// can't hang recall indefinitely; set 0 to restore an unbounded query embed.
	RecallEmbedTimeout time.Duration `env:"MEMINI_RECALL_EMBED_TIMEOUT" envDefault:"2s"`
	// RecallRewriteTimeout bounds the LLM query-expansion call on query_rewrite
	// recalls; past it, recall proceeds with the original query alone rather
	// than riding along the LLM client's much longer HTTP timeout. Default 3s;
	// set 0 to restore an unbounded rewrite call.
	RecallRewriteTimeout time.Duration `env:"MEMINI_RECALL_REWRITE_TIMEOUT" envDefault:"3s"`
	// WriteEmbedTimeout bounds the content embed on the remember path; past it, or on
	// embed error, the memory is stored without a vector (keyword-searchable) and
	// marked pending_embed for background backfill. 0 restores fail-fast writes.
	WriteEmbedTimeout time.Duration `env:"MEMINI_WRITE_EMBED_TIMEOUT" envDefault:"5s"`
	// RecallMinScore is the fused-score floor: candidates below it are dropped
	// before ranking. The default (0.1) is the benched value; it is exposed so a
	// deployment on a different embedder can raise it to trim loosely-relevant
	// injection. Only meaningful with score fusion.
	RecallMinScore float64 `env:"MEMINI_RECALL_MIN_SCORE" envDefault:"0.1"`
	// RecallSemanticReserve reserves up to N of the recall slots for durable
	// tiers (semantic/procedural) so consolidated knowledge is not crowded out by
	// episodic chatter. Exposed because it changes recall composition per
	// deployment: set 0 for pure-relevance recall (no forced durable slots).
	// Reserved slots are relevance-gated — a durable memory is only promoted in
	// when it is relevance-competitive with the entry it displaces.
	RecallSemanticReserve int `env:"MEMINI_RECALL_SEMANTIC_RESERVE" envDefault:"2"`
	// StabilityK is the spaced-repetition strength (Ebbinghaus stability): a
	// short-term memory's effective recall half-life stretches with reinforcement
	// as halfLife*(1+StabilityK*ln(1+access_count)), so a frequently-recalled
	// memory decays more slowly, improving recall of reinforced-but-aged facts
	// (see bench/reinforcement_test.go). Default 1; set 0 to disable (fixed
	// half-life). Only affects short-term tiers with access_count > 0 — durable
	// tiers and never-recalled memories are unchanged.
	StabilityK float64 `env:"MEMINI_STABILITY_K" envDefault:"1"`
	// TurnEchoWindow is the server-wide temporal exclusion window for
	// freshly-captured episodic turns. A just-captured turn
	// (metadata.format="turn") younger than this is dropped from recall by
	// default — a just-captured turn is live context, not long-term memory,
	// and echoing it back makes the agent parrot itself. Callers opt out per
	// call via include_fresh_turns. Default 5m; zero disables it server-wide.
	TurnEchoWindow time.Duration `env:"MEMINI_TURN_ECHO_WINDOW" envDefault:"5m"`

	// EpisodicMinChars drops an episodic write whose substantive content (role
	// scaffolding stripped) is below this many characters — the low-signal
	// per-turn chatter ("keep going", "ok", "hello") that otherwise dominates
	// episodic memory. Only episodic is gated. Default 120 (on); set 0 to disable.
	EpisodicMinChars int `env:"MEMINI_EPISODIC_MIN_CHARS" envDefault:"120"`

	// DistillBatchTokens batches distill-on-write per session: captures
	// accumulate until roughly this many (estimated) tokens, then distill as
	// one LLM call with cross-turn context. 0 restores per-capture distill.
	// Only applies with an LLM configured and to captures with a session_id.
	DistillBatchTokens int `env:"MEMINI_DISTILL_BATCH_TOKENS" envDefault:"1024"`
	// DistillBatchMaxAge flushes a session's buffered captures once the oldest
	// has waited this long, so a quiet session still distills promptly.
	DistillBatchMaxAge time.Duration `env:"MEMINI_DISTILL_BATCH_MAX_AGE" envDefault:"10m"`

	// Consolidation tuning.
	// ConsolidateMode is "async" (default), "sync", or "off".
	ConsolidateMode string `env:"MEMINI_CONSOLIDATE_MODE" envDefault:"async"`
	// ConsolidateMinScore gates the LLM: it runs only when the nearest candidate
	// scores at least this. 0 disables the gate.
	ConsolidateMinScore float64 `env:"MEMINI_CONSOLIDATE_MIN_SCORE" envDefault:"0.3"`

	// Promotion (episodic→semantic distillation). Uses the LLM when configured,
	// the marker extractor otherwise, so it also runs LLM-less.
	// PromoteInterval is how often the promoter runs; 0 disables it.
	PromoteInterval time.Duration `env:"MEMINI_PROMOTE_INTERVAL" envDefault:"24h"`
	// PromoteMinAccess is the minimum access_count for an episodic memory to be
	// considered for promotion.
	PromoteMinAccess int `env:"MEMINI_PROMOTE_MIN_ACCESS" envDefault:"3"`

	// BackfillInterval is how often the vector backfill loop re-embeds
	// memories left vectorless by a degraded write (metadata pending_embed);
	// 0 disables it.
	BackfillInterval time.Duration `env:"MEMINI_BACKFILL_INTERVAL" envDefault:"1m"`

	// SweepInterval is how often the decay sweeper purges expired memories.
	SweepInterval time.Duration `env:"MEMINI_SWEEP_INTERVAL" envDefault:"1h"`
	// ShortTermCap bounds short-term (working+episodic) memories per namespace;
	// the sweeper evicts the lowest-retention ones over the cap. 0 disables it.
	ShortTermCap int `env:"MEMINI_SHORT_TERM_CAP" envDefault:"1000"`
	// TombstoneTTL hard-deletes superseded (deduped/contradicted) memories last
	// updated before now-TTL, reclaiming space. Off by default: tombstones are
	// excluded from recall regardless, so GC is purely a space optimization and
	// removing it is the only irreversible maintenance action. Set e.g. 720h
	// (30d) to enable.
	TombstoneTTL time.Duration `env:"MEMINI_TOMBSTONE_TTL" envDefault:"0"`

	// ActivityLog records reads (recall/get/briefing) and writes
	// (remember/update/forget/supersede) to the activity log that backs the UI's
	// Activity page. Writes are best-effort and off the request path, so the cost
	// is storage, not latency; set false to record nothing.
	ActivityLog bool `env:"MEMINI_ACTIVITY_LOG" envDefault:"true"`
	// ActivityRetention drops activity events older than this. 0 keeps them
	// forever (bounded only by ActivityMaxRows).
	ActivityRetention time.Duration `env:"MEMINI_ACTIVITY_RETENTION" envDefault:"720h"`
	// ActivityMaxRows caps the activity log, dropping the oldest rows beyond it.
	// A busy agent writes several rows per recall, so the cap — not the retention
	// window — is usually what bounds the table. 0 disables the cap.
	ActivityMaxRows int `env:"MEMINI_ACTIVITY_MAX_ROWS" envDefault:"100000"`
	// DemoteAfter demotes durable memories older than this to the episodic tier
	// when they have never been recalled, are not important, and are
	// uncorroborated (low confidence) — so an old bulk import ages out while
	// facts the agent actually uses or establishes are kept. Default 168h (7d);
	// set to 0 to disable.
	DemoteAfter time.Duration `env:"MEMINI_DEMOTE_AFTER" envDefault:"168h"`

	// Dedup tuning. The dedup pass collapses near-duplicate memories
	// (embedding similarity ≥ DedupSimilarity) into a single representative
	// per cluster; the rest are tombstoned (SupersededBy → representative),
	// not hard-deleted, so the action is reversible. Exposed on-demand via
	// POST /v1/dedup and run as a periodic store-wide background job every
	// DedupInterval (daily by default, so a store stays clean with no manual
	// intervention). Set MEMINI_DEDUP_INTERVAL=0 to disable the periodic pass.
	DedupInterval time.Duration `env:"MEMINI_DEDUP_INTERVAL" envDefault:"24h"`
	// DedupSimilarity is the embedding-similarity threshold for two memories to
	// count as members of the same near-duplicate cluster. Higher is stricter,
	// so raise it if the pass is collapsing memories that were only superficially
	// alike, and lower it if obvious restatements survive.
	DedupSimilarity float64 `env:"MEMINI_DEDUP_SIMILARITY" envDefault:"0.85"`
	// DedupTiers is an optional comma-separated list restricting the periodic
	// pass to specific tiers (working,episodic,semantic,procedural). Empty
	// means all tiers.
	DedupTiers string `env:"MEMINI_DEDUP_TIERS" envDefault:""`
	// DedupLLMMerge (opt-in, default off) enables LLM-based content merging
	// during the periodic dedup pass. Each cluster's content is merged into a
	// single comprehensive memory before tombstoning duplicates. Requires an
	// LLM (MEMINI_LLM_BASE_URL); when false or no LLM, the representative
	// keeps its original content. Defaults off to preserve existing behavior.
	DedupLLMMerge bool `env:"MEMINI_DEDUP_LLM_MERGE" envDefault:"false"`

	// UIEnabled mounts the embedded admin UI at /. Enabled by default; set
	// MEMINI_UI_ENABLED=false to run a headless API/MCP-only service.
	UIEnabled bool `env:"MEMINI_UI_ENABLED" envDefault:"true"`

	// APIKey is the admin, or bootstrap, credential. When set, every request
	// must present it as a bearer token, it also gates /metrics, and it is the
	// only credential allowed to manage other keys through /v1/keys. Named keys
	// (see APIKeysFile and `memini key`) are the per-person credentials; this is
	// the one that creates them.
	//
	// Note that when this is set and the admin UI is enabled, the server embeds
	// the key in the UI shell so the same-origin UI can authenticate without the
	// operator pasting it. Anyone who can load / can therefore read the key.
	// Expose the UI only where reaching it already implies trust, move it to its
	// own listener with UIAddr, or turn it off with UIEnabled.
	APIKey string `env:"MEMINI_API_KEY"`

	// APIKeysFile (optional; K2b), when set, names a YAML file of
	// declaratively managed API keys — the GitOps-friendly counterpart to
	// the api_keys table (which is managed imperatively via `memini key
	// ...` / a future /v1/keys API, K3b). The file is loaded exactly ONCE at
	// boot (see internal/apiauth.LoadFileKeys); there is no live reload
	// today — a GitOps rollout restarts the pod on every change to the
	// file's content, which already picks up edits, so a SIGHUP-triggered
	// in-process reload is a reasonable future addition but is not built
	// here. Absent (the default) is a complete no-op: zero behavior change
	// versus a server built before this field existed.
	//
	// Format (see internal/apiauth/testdata/api_keys.example.yaml for a
	// runnable example):
	//
	//	keys:
	//	  - name: alex                             # required, unique within the file
	//	    hash: "<hex sha-256 of the secret>"     # exactly one of hash|secret
	//	    home: personal/alex                     # optional
	//	    default_namespace: acme                 # optional
	//	    disabled: false                         # optional, default false
	//	  - name: ci
	//	    secret: "plaintext secret"              # allowed: the file itself is
	//	                                             # the secret store (e.g.
	//	                                             # SOPS-encrypted at rest);
	//	                                             # hashed at load, never kept
	//	                                             # in memory as plaintext
	//
	// Boot validation is fail-loud: malformed YAML, a missing name, both or
	// neither of hash/secret, a hash that isn't valid hex-encoded SHA-256, a
	// duplicate name within the file, or an invalid home/default_namespace
	// (httputil.ValidateNamespace, after httputil.NormalizeNamespace) all
	// refuse the boot with a message naming this file and the offending
	// entry. A file key that shares a name with an existing api_keys table
	// row wins at auth time (internal/apiauth.Config.Authenticate: file
	// checked before the table); the server logs a warning at boot listing
	// which table keys are shadowed this way.
	APIKeysFile string `env:"MEMINI_API_KEYS_FILE"`

	// Multi-tenancy. The fallback namespace when no header is sent; the header
	// name itself is fixed (DefaultNamespaceHeader).
	DefaultNamespace string
	NamespaceSrc     NamespaceSource

	// Home is the caller's personal namespace: merged read-only (durable
	// tiers only) into the default read set on every recall/briefing/answer,
	// on top of the request namespace and its ancestors. Client-side only —
	// the server never derives it. On HTTP transports it is carried per-request
	// by the X-Memini-Home header (DefaultHomeHeader); this env var is what the
	// stdio MCP server (`memini mcp`) resolves instead, since stdio has no
	// headers. Empty means no home leg (unset by default).
	Home string `env:"MEMINI_HOME"`
}

Config is the fully-resolved runtime configuration. Environment-backed fields are parsed by github.com/caarlos0/env via their `env` tags; an absent variable falls back to `envDefault`, while a set-but-empty variable is taken verbatim. DefaultNamespace/NamespaceSrc are resolved separately (see resolveDefaultNamespace) and carry no tag.

func Load

func Load() (*Config, error)

Load reads configuration from the environment and validates it.

func (*Config) DedupTierList added in v0.0.8

func (c *Config) DedupTierList() []memory.Tier

DedupTierList parses MEMINI_DEDUP_TIERS into the tiers the periodic dedup pass is restricted to. Empty/unset returns nil, meaning all tiers. Values are validated in validate(), so the result is safe to use directly.

func (*Config) LLMEnabled

func (c *Config) LLMEnabled() bool

LLMEnabled reports whether the opt-in LLM pipeline is configured.

func (*Config) RerankEnabled added in v0.0.4

func (c *Config) RerankEnabled() bool

RerankEnabled reports whether recall reranking is configured.

func (*Config) RerankIsLLM added in v0.0.4

func (c *Config) RerankIsLLM() bool

RerankIsLLM reports whether reranking uses the chat LLM rather than a cross-encoder URL.

type NamespaceSource

type NamespaceSource string

NamespaceSource records how DefaultNamespace was resolved, useful for startup logging and debug surfaces.

const (
	NamespaceFromEnv     NamespaceSource = "env"      // MEMINI_DEFAULT_NAMESPACE / MEMINI_NAMESPACE
	NamespaceFromGit     NamespaceSource = "git"      // git rev-parse --show-toplevel basename
	NamespaceFromCWD     NamespaceSource = "cwd"      // filepath.Base(cwd)
	NamespaceFromLiteral NamespaceSource = "fallback" // literal "default"
)
const NamespaceFromGitRemote NamespaceSource = "git-remote"

NamespaceFromGitRemote marks a namespace resolved from the `git remote get-url origin` repo name — the order the plugins use.

const NamespaceFromOverride NamespaceSource = "override"

NamespaceFromOverride marks a namespace that came from an explicit per-project override the user set (via `/memini:namespace` or an equivalent command in another harness), rather than from the environment or git.

func ResolveDirNamespace added in v0.0.11

func ResolveDirNamespace(dir string) (string, NamespaceSource)

ResolveDirNamespace resolves a directory's namespace from git, ignoring any MEMINI_NAMESPACE env override: git remote origin repo name, then the worktree basename, then the directory basename. The claude-code backfill uses this so each project's transcripts land in their own namespace instead of collapsing into one global env namespace — which is exactly the pooling failure to avoid.

func ResolvePluginNamespace added in v0.0.11

func ResolvePluginNamespace(dir string) (string, NamespaceSource)

ResolvePluginNamespace resolves a namespace the way the Claude Code / OpenClaw plugins do (plugin/scripts/_shared.mjs resolveProject):

  1. the per-project override (~/.config/memini/overrides.json), if set
  2. MEMINI_NAMESPACE (or MEMINI_DEFAULT_NAMESPACE) env, if non-empty
  3. repo name from `git remote get-url origin` (stable across worktrees/clones)
  4. basename of `git rev-parse --show-toplevel`
  5. basename of dir

then nested under a per-agent segment when MEMINI_AGENT is set (step 6, mirroring _shared.mjs withAgent). It differs from resolveDefaultNamespace (the server's header-less fallback) in step 3 (the server skips the git-remote step) and step 6; `memini doctor` uses both to flag the divergence that lands writes where recall doesn't look. The server's resolution is intentionally left unchanged so existing stores keyed by the worktree basename are not silently relocated.

The override outranks the env var, matching the client exactly. Getting this backwards would be worse than not reading the file at all: doctor would report a namespace the plugin does not actually use, on precisely the machines where a globally exported MEMINI_NAMESPACE is the problem being diagnosed.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL