deja

Persistent memory for AI agents. Agents learn from runs — deja remembers across them.

Three ways to use deja

	deja-local	deja-edge	deja (hosted)
What it is	In-process SQLite memory	Cloudflare Durable Object memory	Hosted Cloudflare Worker service
Search	Vector embeddings (all-MiniLM-L6-v2)	FTS5 full-text search	Vectorize + Workers AI embeddings
Runtime	Bun	Cloudflare Workers	Cloudflare Workers
Dependencies	None (runs locally)	None (runs in your DO)	Vectorize index + Workers AI
Latency	Zero (in-process)	Zero (in-DO)	Network round-trip
Install	npm install deja-local	npm install deja-edge	Clone + wrangler deploy
Best for	Local agents, scripts, CLI tools	Edge agents running in Workers	Multi-tenant, team-shared memory

deja-local

Memory lives in a local SQLite file. Vector search via ONNX embeddings on CPU. No network, no API keys. Requires Bun.

import { createMemory } from 'deja-local'

const mem = createMemory({ path: './agent.db' })
await mem.remember('always run migrations before deploying')
const results = await mem.recall('deploying to staging')

Full docs →

deja-edge

Memory lives in a Cloudflare Durable Object's SQLite. FTS5 full-text search with BM25 ranking. No Vectorize, no Workers AI — just text matching. Zero external dependencies.

import { createEdgeMemory } from 'deja-edge'

export class MyDO extends DurableObject {
  private memory = createEdgeMemory(this.ctx)

  async onRemember(text: string) {
    return this.memory.remember(text)
  }
  async onRecall(context: string) {
    return this.memory.recall(context)
  }
}

Or use the drop-in DO class with HTTP routes:

export { DejaEdgeDO } from 'deja-edge/do'

Full docs →

deja (hosted)

Full-featured hosted service with Vectorize semantic search, scoped memory, working state, secrets, and MCP support. Deploy your own instance or use as part of filepath.

git clone https://github.com/acoyfellow/deja && cd deja
bun install && wrangler login
wrangler vectorize create deja-embeddings --dimensions 384 --metric cosine
wrangler secret put API_KEY
bun run deploy

Connect via REST, the client package, or MCP:

{
  "mcpServers": {
    "deja": {
      "type": "http",
      "url": "https://your-deja-instance.workers.dev/mcp",
      "headers": { "Authorization": "Bearer ${DEJA_API_KEY}" }
    }
  }
}

Core concepts

All three systems share the same mental model:

• Remember — store a memory (text, optionally with trigger/context/confidence)
• Recall — search for relevant memories given a context
• Confirm / Reject — feedback loop. Confirmed memories rise, rejected ones fade
• Forget — permanently delete a memory

Memories are deduplicated at write time. Conflicting memories (same topic, different content) are automatically resolved — the newer one supersedes the older one.

Time-based confidence decay

Stale memories don't sit at 0.5 forever. At recall time, confidence decays exponentially based on how recently the memory was created or last recalled:

decayedConfidence = storedConfidence × 0.5^(daysSince / 90)

A memory untouched for 90 days has its effective confidence halved. Recalling a memory resets its decay clock — actively used knowledge stays fresh. Stored confidence is never mutated by decay; only confirm() and reject() change the stored value.

Agent attribution

Track which agent stored a memory with the optional source parameter:

await mem.remember('always use pnpm', { source: 'deploy-agent' })

Anti-patterns (deja-local & deja-edge)

When a memory is rejected enough that its confidence drops below 0.15, it auto-inverts into an anti-pattern — a warning that actively surfaces during recall:

Before: "use eval for JSON parsing"  (confidence: 0.05, type: "memory")
After:  "KNOWN PITFALL: use eval for JSON parsing"  (confidence: 0.5, type: "anti-pattern")

Negative knowledge is as valuable as positive knowledge. Anti-patterns participate in recall normally and warn agents away from known mistakes.

Hosted service API

The hosted service adds features beyond basic memory:

• Scoped memory — shared, agent:<id>, session:<id>, or custom scopes
• Working state — live snapshots + event streams for in-progress work
• Secrets — scoped key-value storage
• Loop runs — track optimization loops with auto-learning from outcomes

REST endpoints: /learn, /inject, /query, /learnings, /stats, /state/:runId, /secret, /run

Full reference: https://deja.coey.dev/docs

Architecture

• deja-local: Bun SQLite + ONNX embeddings (all-MiniLM-L6-v2). In-memory vector index loaded at startup.
• deja-edge: Cloudflare DO SQLite + FTS5. Porter stemming tokenizer. BM25 ranking blended with confidence.
• deja (hosted): Cloudflare Worker + Durable Object + Vectorize + Workers AI. Per-user isolation via API key.

License

MIT