Skip to content

README.md

Latest commit

 

History

History
102 lines (74 loc) · 4.02 KB

README.md

File metadata and controls

102 lines (74 loc) · 4.02 KB
Raw

@dawn-example/memory — long-term memory, backend-switchable

A one-route Dawn app (notes) with a note-taking agent that has durable, cross-session memory. It ships with a zero-setup SQLite backend and switches to Postgres + pgvector with a single environment variable — the same app code, a different store.

The route

  • src/app/notes/index.ts — the agent (gpt-5-mini) with a remember/recall system prompt.
  • src/app/notes/memory.ts — a semantic memory schema (subject / predicate / value), route-scoped.

The remember and recall tools are generated from memory.ts.

Backends

The backend is chosen at load time from the environment (see dawn.config.ts):

Env Store Recall
(none) SQLite (default) keyword-only
OPENAI_API_KEY SQLite hybrid keyword + vector
DATABASE_URL Postgres/pgvector keyword-only
DATABASE_URL + OPENAI_API_KEY Postgres/pgvector hybrid keyword + vector

The two toggles are independent. DATABASE_URL swaps the store; OPENAI_API_KEY lights up vector/semantic recall (text-embedding-3-small, 1536 dims). Both the store and the embedder connect lazily, so nothing touches the network until the first remember/recall.

Run it (SQLite, zero setup)

pnpm --filter @dawn-example/memory dev

Memory persists to .dawn/memory.sqlite. That's it — no key, no database.

Run it against Postgres + pgvector

Start a pgvector-enabled Postgres:

docker run --rm -e POSTGRES_PASSWORD=postgres -p 5432:5432 pgvector/pgvector:pg16

Point the app at it (and, optionally, add a key for vector recall):

export DATABASE_URL="postgres://postgres:postgres@localhost:5432/postgres"
export OPENAI_API_KEY="sk-..."   # optional — enables hybrid keyword+vector recall
pnpm --filter @dawn-example/memory dev

The app creates its tables + HNSW index on first write.

Distillation

Once the app has accumulated some history, the two distillation passes compact it — consolidation summarizes old episodes per (namespace, ISO week); reflection derives durable insights:

# See the plan without spending a token (no model call, no key needed):
pnpm --filter @dawn-example/memory exec dawn memory consolidate --dry-run
pnpm --filter @dawn-example/memory exec dawn memory reflect --dry-run

# Run them for real (needs OPENAI_API_KEY — the default model is gpt-5-mini):
pnpm --filter @dawn-example/memory exec dawn memory consolidate
pnpm --filter @dawn-example/memory exec dawn memory reflect

Both are threshold-aware no-ops, so running them before there is anything to distill just prints nothing to consolidate / nothing to reflect on and exits 0 without constructing a model. Reflection writes its insights as candidates by default — review them with dawn memory list and promote with dawn memory approve <id>. See the distillation docs for every flag and the memory.distill config block.

Continuous dogfood

packages/testing/test/memory-example-dogfood.test.ts drives this real example app through a scripted remember → recall flow:

  • Always (CI-safe, no key, no Docker): the default SQLite backend, proving the memory route works end-to-end.
  • Gated (DAWN_TEST_PGVECTOR=1, Docker): the same flow against a Testcontainers Postgres, proving recall works through pgvector.
# CI-safe (SQLite) — gated block auto-skips:
pnpm --filter @dawn-ai/testing exec vitest run test/memory-example-dogfood.test.ts

# Local hands-on pgvector dogfood (needs Docker):
DAWN_TEST_PGVECTOR=1 pnpm --filter @dawn-ai/testing exec vitest run test/memory-example-dogfood.test.ts