agent_with_memory/

Agent with memory (agent_with_memory)

Examples that wire long-term memory into agent-sdk-go. Pick one backend per run.

Store mode is selected with MEMORY_STORE_MODE (always or ondemand; default ondemand). Both backends support both modes. task examples:local runs four combinations (each backend × each mode).

Uses the same Docker stack as retriever examples (../docker/). task infra:weaviate:up / task infra:pgvector:up creates the memory class/table (AgentMemory / agent_memories) in addition to retriever schema. No seed rows for memory — rows are written by agent runs.

Prerequisites

• Runtime — AGENT_RUNTIME=local (default): in-process, no Temporal. Optional AGENT_RUNTIME=temporal: from examples/, run task infra:temporal:up (and task infra:temporal:wait if the example fails to connect). See temporal-setup.md.
• examples/.env — LLM_APIKEY, LLM_MODEL, and EMBEDDING_OPENAI_APIKEY (see .env.defaults)
• Task (go-task) and Docker (task infra:weaviate:up or task infra:pgvector:up)

From examples/:

task infra:status    # see what is up

Example behavior

• Store mode — MEMORY_STORE_MODE=always extracts and stores at run end; ondemand registers save_memory for the LLM during the run (default).
• No CLI args — two runs in one process: run 1 stores a preference, run 2 recalls it.
• With args — single custom prompt.
• Scope — MEMORY_USER_ID in .env (default demo-user); must be the same across runs you want to share memories.

Set MEMORY_RECALL_ENABLED=false in .env for store-only (skip load before LLM).

Use SHOW_TELEMETRY=true to see total_memory_stores on run 1 when store succeeds.

Weaviate

Weaviate embeds memory text via nearText (text2vec-openai in Docker).

Setup

cd examples
task infra:weaviate:up
task infra:weaviate:down   # when finished

Compose: docker/docker-compose.yml. Seed: docker/weaviate/seed.sh (creates class AgentMemory).

EMBEDDING_OPENAI_APIKEY must be set in examples/.env before up. After a key change: task infra:weaviate:down && task infra:weaviate:up.

Verify the memory class exists:

curl -s http://localhost:8080/v1/schema | jq '.classes[].class'
# expect Document and AgentMemory

Environment

WEAVIATE_HOST=localhost:8080
WEAVIATE_SCHEME=http
WEAVIATE_MEMORY_CLASS=AgentMemory
MEMORY_USER_ID=demo-user
MEMORY_RECALL_ENABLED=true
MEMORY_RECALL_LIMIT=10
MEMORY_RECALL_MIN_SCORE=0.35

Run

go run ./agent_with_memory/weaviate
go run ./agent_with_memory/weaviate "Remember my favorite color is blue"

SHOW_TELEMETRY=true go run ./agent_with_memory/weaviate

Weaviate troubleshooting

LOG_LEVEL=debug go run ./agent_with_memory/weaviate

pgvector

Client-side OpenAI-compatible embeddings, then cosine search in Postgres (pgvector).

Setup

cd examples
task infra:pgvector:up
task infra:pgvector:down   # when finished

Schema: docker/pgvector/setup.sql (table agent_memories). Seed: docker/pgvector/seed.sh.

Default DSN (in .env.defaults): postgres://postgres:secret@localhost:5432/vectordb?sslmode=disable

Verify the memory table exists:

docker exec pgvector psql -U postgres -d vectordb -c "\d agent_memories"

Environment

PGVECTOR_DSN=postgres://postgres:secret@localhost:5432/vectordb?sslmode=disable
PGVECTOR_MEMORY_TABLE=agent_memories
EMBEDDING_OPENAI_MODEL=text-embedding-3-small
EMBEDDING_OPENAI_APIKEY=sk-...
MEMORY_USER_ID=demo-user
MEMORY_RECALL_ENABLED=true
MEMORY_RECALL_LIMIT=10
MEMORY_RECALL_MIN_SCORE=0.35

With Anthropic/Gemini chat, EMBEDDING_OPENAI_APIKEY is still required (not LLM_APIKEY).

Run

go run ./agent_with_memory/pgvector
go run ./agent_with_memory/pgvector "What answer style do I prefer?"

SHOW_TELEMETRY=true go run ./agent_with_memory/pgvector

pgvector troubleshooting

LOG_LEVEL=debug go run ./agent_with_memory/pgvector