Documentation
¶
Overview ¶
Package operations holds LadyM's cognitive operations (activation, recall, consolidation, decay, proceduralization, supersedes, attention, L5/L6).
Index ¶
- func ActivationScore(mem *schema.Memory, querySimilarity float64, w config.ActivationWeights, ...) float64
- func FrequencyFactor(accessCount int) float64
- func GraphFactor(mem *schema.Memory, neighbourCounts map[string]int, weight float64) float64
- func InferQueryTypes(query string) []schema.MemoryType
- func IsConsolidated(m *schema.Memory) bool
- func IsRetired(mem *schema.Memory) bool
- func LatestInChain(store storage.Store, memID string) (string, error)
- func NeighbourCountsFor(mems []*schema.Memory, counts map[string]int) map[string]int
- func Recall(store storage.Store, embedder storage.EmbeddingProvider, query string, ...) (*schema.RecallResponse, error)
- func RecencyFactor(lastAccessAt, halfLifeS, now float64) float64
- func Retire(store storage.Store, old *schema.Memory, newID string) error
- func TypeBoostForQuery(mem *schema.Memory, queryTypes []schema.MemoryType, weight float64) float64
- type Action
- type ConsolidationReport
- type DecayReport
- type GateDecision
- type L5ExtractionReport
- type L6PredictionReport
- type LLMClassifier
- type ProceduralizeReport
- type System2Report
- type System2Runner
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func ActivationScore ¶
func ActivationScore(mem *schema.Memory, querySimilarity float64, w config.ActivationWeights, neighbourCounts map[string]int, queryTypes []schema.MemoryType, now float64) float64
ActivationScore computes the full ACT-R-inspired activation score.
func FrequencyFactor ¶
FrequencyFactor is the diminishing-returns log curve log(1 + n).
func GraphFactor ¶
GraphFactor is spreading activation: items with more current graph neighbours get a small boost.
func InferQueryTypes ¶
func InferQueryTypes(query string) []schema.MemoryType
InferQueryTypes heuristically detects whether a query is about code (boosting code_symbol/code_file) or a how-to (boosting playbook).
func IsConsolidated ¶ added in v0.5.2
IsConsolidated reports whether the episode has completed a consolidation pass (stamped for every ADD/UPDATE/DELETE/NOOP verdict).
func IsRetired ¶
IsRetired reports whether mem was retired by an UPDATE or DELETE consolidation pass. A nil memory is treated as not-retired.
func LatestInChain ¶
LatestInChain walks supersedes edges forward and returns the newest version id.
func NeighbourCountsFor ¶
NeighbourCountsFor projects a {id: count} map onto just the given memories.
func Recall ¶
func Recall(store storage.Store, embedder storage.EmbeddingProvider, query string, cfg *config.Config, workspace string, topK int, layers []schema.Layer, types []schema.MemoryType, minSimilarity float64) (*schema.RecallResponse, error)
Recall runs the full two-tier retrieval pipeline.
func RecencyFactor ¶
RecencyFactor is exponential decay: 1.0 right after access, 0.5 after halfLifeS seconds.
func Retire ¶
Retire retires old. When newID is non-empty it writes a supersedes edge old→new (UPDATE chain); otherwise it sets superseded=true (DELETE). Outgoing still-valid edges of old are closed so the graph does not leak through a retired node.
func TypeBoostForQuery ¶
TypeBoostForQuery boosts items whose type matches a query-type prior.
Types ¶
type Action ¶
type Action string
Action is one of the four consolidation decisions (mem0's ADD/UPDATE/DELETE/NOOP).
type ConsolidationReport ¶
type ConsolidationReport struct {
Actions map[string]int
KeptEpisodes int
PromotedToSemantic int
SkippedConsolidated int
Details []map[string]any
}
ConsolidationReport reports the outcome of a consolidation pass.
func Consolidate ¶
func Consolidate(store storage.Store, embedder storage.EmbeddingProvider, cfg *config.Config, workspace string, llmClassify LLMClassifier, since float64) (*ConsolidationReport, error)
Consolidate promotes salient episodic events into semantic facts. Selection covers the workspace's unconsolidated episodes (see consolidatedAtKey); when since is non-zero it additionally bounds the scan to episodes created at or after that time, so 0 means "everything still pending".
type DecayReport ¶
DecayReport reports the outcome of a decay pass.
func Decay ¶
func Decay(store storage.Store, workspace string, weights *config.ActivationWeights, maxAgeS, activationFloor, now float64, dryRun bool) (*DecayReport, error)
Decay forgets episodic events whose activation has fallen below the floor. Code analysis, playbooks, and edges are never auto-forgotten.
type GateDecision ¶
type GateDecision struct {
Action string // "pass" | "rewrite" | "drop"
Content string // populated only on rewrite
Reason string
}
GateDecision is the outcome of the attention gate.
func AttentionGate ¶
func AttentionGate(content string, cfg *config.Config, store storage.Store, getAgent func(string) (providers.LLMProvider, error), layer schema.Layer) (GateDecision, error)
AttentionGate applies the pre-remember filter to content destined for layer. getAgent resolves the LLM agent bound to "attention_gate" (nil for offline).
type L5ExtractionReport ¶
type L5ExtractionReport struct {
NewModels int
MergedModels int
Clusters []map[string]any
Skipped bool
}
L5ExtractionReport reports the outcome of L5 mental-model extraction.
func ExtractL5 ¶
func ExtractL5(store storage.Store, embedder storage.EmbeddingProvider, cfg *config.Config, workspace string, llm providers.LLMProvider, prompt string) (*L5ExtractionReport, error)
ExtractL5 clusters uncovered L2/L3 memories into mental models.
type L6PredictionReport ¶
type L6PredictionReport struct {
Predictions int
ExpiredRetired int
EpisodesSeen int
WatermarkUpdatedTo float64
Details []map[string]any
Skipped bool
}
L6PredictionReport reports the outcome of L6 forward-intent prediction.
func PredictL6 ¶
func PredictL6(store storage.Store, embedder storage.EmbeddingProvider, cfg *config.Config, workspace string, llm providers.LLMProvider, prompt string) (*L6PredictionReport, error)
PredictL6 predicts next intents from recent episodes, with TTL expiry.
type LLMClassifier ¶
LLMClassifier is a pluggable (candidate, similar) → (Action, newText) classifier. A non-nil error aborts the consolidation pass (Python: exceptions propagate).
type ProceduralizeReport ¶
type ProceduralizeReport struct {
ClustersExamined int
PlaybooksCreated int
Actions map[string]int
Details []map[string]any
}
ProceduralizeReport reports the outcome of a proceduralization pass.
func Proceduralize ¶
func Proceduralize(store storage.Store, embedder storage.EmbeddingProvider, cfg *config.Config, workspace string, minClusterSize int, similarityThreshold float64) (*ProceduralizeReport, error)
Proceduralize clusters successful episodic events into L3 playbooks.
type System2Report ¶
type System2Report struct {
Consolidate any
Proceduralize any
L5 any
L6 any
Decay any
SkippedLLMSteps bool
}
System2Report is the outcome of one background consolidation cycle.
func RunSystem2Cycle ¶
func RunSystem2Cycle(runner System2Runner, workspace string) (*System2Report, error)
RunSystem2Cycle runs one System 2 cycle through the runner.
type System2Runner ¶
type System2Runner interface {
Consolidate(workspace string, since float64) (*ConsolidationReport, error)
Proceduralize(workspace string, minClusterSize int) (*ProceduralizeReport, error)
ExtractMentalModels(workspace string) (*L5ExtractionReport, error)
PredictForwardIntents(workspace string) (*L6PredictionReport, error)
Decay(workspace string, dryRun bool, maxAgeS, activationFloor float64) (*DecayReport, error)
CountRecentEpisodes(workspace string) (int, error)
MinEpisodesToRun() int
}
System2Runner is the interface the engine satisfies to run a System 2 cycle (kept as an interface so operations does not import engine).
Source Files