memex

A production-grade persistent memory service for AI agents. Agents forget everything between sessions by default — memex fixes that. It stores, retrieves, and ranks conversation memory using semantic search with recency decay, so agents surface what's relevant and recent, not just what's semantically closest.

POST /v1/memories          → store a memory, embed it, persist to Postgres
POST /v1/memories/search   → retrieve top-k memories ranked by similarity + recency
DELETE /v1/memories/{id}   → forget a specific memory
GET  /v1/memories/count    → how many memories does this agent/user have
GET  /health               → liveness + DB connectivity check
GET  /metrics              → Prometheus metrics

Architecture

caller (agent / app)
        │
        ▼
  FastAPI (async)
        │
   ┌────┴────┐
   │         │
embeddings  asyncpg pool (min=5, max=20)
(fastembed  │
 ONNX,      ▼
 local)  PostgreSQL 16
           pgvector extension
           ivfflat index (cosine)

Write path: content → fastembed ONNX inference (local, ~12 ms CPU, BAAI/bge-small-en-v1.5) → INSERT with 384-dim vector → return memory ID.

Read path: query → embed → pgvector cosine search (top_k × 3 candidates) → re-rank with recency decay in Python → return top_k results with scores.

Design decisions

1. Recency decay on top of semantic search

Pure vector similarity returns the most semantically similar memories, not the most useful ones. A fact from 90 days ago that's a 0.95 similarity match is often less useful than a 0.80 match from yesterday.

Score formula:

score = α × cosine_similarity + (1 − α) × exp(−λ × age_days)

Where λ = ln(2) / half_life_days (default: 30 days, so a 30-day-old memory has 50% recency weight).

α is configurable per request (default 0.7). Task-focused agents use higher α (semantic dominates). Conversational agents use lower α (recency matters more).

2. Fetch 3× candidates, re-rank in Python

The pgvector query returns top_k × 3 candidates sorted by pure similarity. Python re-ranks with the decay formula and slices to top_k. This prevents recency decay from starving high-similarity older memories — they're still in the candidate pool.

At 10× scale (>1M memories per agent): push the scoring into a Postgres function using pg_proc to eliminate the Python re-ranking round-trip.

3. asyncpg + explicit pool sizing over SQLAlchemy async

SQLAlchemy adds ORM overhead on every query. The hot retrieval path — embed, query, re-rank — needs to be tight. asyncpg gives direct control over pool min/max (same instinct as tuning HikariCP in Java). pgvector queries require raw SQL for the <=> operator anyway.

Pool defaults: min=5, max=20. Right-size for a single-instance deployment. Override via DB_MAX_POOL_SIZE env var.

4. Rate limiting in Postgres, not Redis

Sliding window counter via upsert. One fewer dependency. Correct under concurrent requests (transactional upsert). At 10× scale with distributed deployments: replace with Redis INCR + EXPIRE — atomic operations, no lock contention.

5. ivfflat index, not HNSW

ivfflat has lower build cost and lower memory footprint — the right tradeoff at small-to-medium scale (<1M vectors). lists=100 works well up to ~1M rows. At 10× scale: switch to HNSW (m=16, ef_construction=64) for better recall at the cost of higher memory and build time.

Running locally

Prerequisites: Docker and Docker Compose. No API keys required — the entire stack runs locally.

git clone https://github.com/ayushagrawal288/memex
cd memex
docker compose up

The API is live at http://localhost:8000. Interactive docs at http://localhost:8000/docs.

API reference

Store a memory

curl -X POST http://localhost:8000/v1/memories \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "my-agent",
    "user_id": "user-123",
    "content": "User prefers concise responses and dislikes verbose explanations.",
    "memory_type": "semantic",
    "importance": 1.2
  }'

{
  "id": "3fa85f64-5717-4562-b3fc-2c963f66afa6",
  "agent_id": "my-agent",
  "user_id": "user-123",
  "content": "User prefers concise responses and dislikes verbose explanations.",
  "importance": 1.2,
  "memory_type": "semantic",
  "created_at": "2026-05-26T10:30:00Z",
  "score": null
}

Search memories

curl -X POST http://localhost:8000/v1/memories/search \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "my-agent",
    "user_id": "user-123",
    "query": "how does this user like to communicate",
    "top_k": 5,
    "alpha": 0.7
  }'

{
  "results": [
    {
      "id": "3fa85f64-...",
      "content": "User prefers concise responses and dislikes verbose explanations.",
      "memory_type": "semantic",
      "created_at": "2026-05-26T10:30:00Z",
      "score": 0.8921
    }
  ],
  "query": "how does this user like to communicate",
  "total": 1
}

Memory types

Load test results

Run on a MacBook M-series, Docker Desktop, single Postgres instance:

locust -f scripts/load_test.py --host=http://localhost:8000 \
       --headless -u 50 -r 10 -t 60s

Realistic load (50 users, 100–300 ms think time — models actual agent traffic):

Saturation test (500 users, minimal think time — finds the throughput ceiling):

Run on MacBook M-series, Docker Desktop (4 CPUs), 4 uvicorn workers, 16 threads/worker.
Embeddings: local ONNX (BAAI/bge-small-en-v1.5) — zero external API calls, zero cost.

Why the ceiling is ~120 RPS:
Every write and every search requires one ONNX inference (~10–15 ms on CPU). With 4 Docker CPUs: 4 cores / 12 ms ≈ 333 embeddings/s theoretical max. After Python overhead, DB queries, and asyncio scheduling: ~120 RPS actual.

Path to higher throughput:

Project structure

memex/
├── app/
│   ├── main.py                  # REST API — FastAPI, lifespan, router registration
│   ├── mcp_server.py            # MCP server — single-worker FastAPI on port 8001
│   ├── core/
│   │   └── config.py            # All settings, loaded from env
│   ├── db/
│   │   └── pool.py              # asyncpg pool, migrations
│   ├── models/
│   │   └── schemas.py           # Pydantic request/response models
│   ├── services/
│   │   ├── embeddings.py        # fastembed ONNX inference (local, zero API calls)
│   │   ├── local_summarizer.py  # Extractive summariser — Jaccard dedup + TF scoring
│   │   ├── memory.py            # Core write/search/scoring logic
│   │   ├── metrics.py           # Prometheus metric definitions
│   │   ├── summarizer.py        # Background summarisation job
│   │   └── rate_limit.py        # Sliding window rate limiter
│   └── api/routes/
│       ├── memories.py          # Memory endpoints
│       ├── health.py            # Health + readiness
│       └── mcp_tools.py         # MCP tool definitions (store, search, delete, count)
├── scripts/
│   └── load_test.py             # Locust load test
├── docker-compose.yml
├── Dockerfile
└── requirements.txt

Observability

docker compose up starts Prometheus and Grafana alongside the API:

The Grafana dashboard is provisioned automatically. Panels:

• HTTP request rate + latency p50/p99 — from prometheus-fastapi-instrumentator
• Embedding API latency p50/p99 — per-attempt histogram by operation (embed / embed_batch)
• Memory operations/s — create, search, delete throughput
• DB pool utilisation — active vs idle connections (update interval: 15 s)
• Summariser activity — memories condensed per hour, run outcomes
• Embedding errors/min — by operation and error type

Custom metrics are in app/services/metrics.py and exposed on /metrics alongside the standard FastAPI instrumentator metrics.

MCP endpoint

memex exposes itself as an MCP server so any MCP-aware agent (Claude Desktop, Claude Code, custom agents) can store and retrieve memories without custom HTTP integration.

Transport: Streamable HTTP (MCP 2024-11-05 spec). Single-worker process on port 8001 — session state is in-process, so a separate service avoids sticky-session complexity while keeping the REST API's multi-worker throughput.

Tools:

Connect from Claude Desktop

Add to ~/.config/claude/claude_desktop_config.json:

{
  "mcpServers": {
    "memex": {
      "type": "streamable-http",
      "url": "http://localhost:8001/mcp/"
    }
  }
}

Connect from Claude Code

claude mcp add --transport http memex http://localhost:8001/mcp/

Design: why a separate service

The MCP Streamable HTTP transport is session-stateful — initialize, tools/list, and tools/call must all reach the same server process. The REST API runs 4 uvicorn workers with round-robin routing; routing different MCP requests to different workers breaks session state.

Running a dedicated single-worker MCP service on port 8001 avoids sticky-session infrastructure (nginx ip_hash, Redis session store) while keeping the REST API fully multi-worker.

Memory summarisation

Runs as a background asyncio task on a configurable interval (default: every 5 minutes). Finds any (agent_id, user_id) pair where episodic memory count exceeds a threshold, condenses the oldest batch into a single semantic memory, then deletes the originals. Fully local — no LLM API calls.

How it summarises: Pure Python extractive algorithm. Sentences are deduplicated by Jaccard similarity (≥ 0.7 threshold), scored by word frequency (TF), and the top-N are returned in original order. ~1 ms per summarisation, zero dependencies beyond the standard library.

Why episodic-only: Episodic memories are conversation events with natural time-based obsolescence. Semantic and procedural memories encode facts and skills — silently condensing them risks precision loss; they age out via recency decay instead.

Concurrency safety: Uses pg_try_advisory_xact_lock keyed on hashtext(agent_id|user_id). The lock is held only during the DB write transaction, not during the embedding call.

Tune via env vars:

What's next

• Memory summarisation — background job to condense old episodic memories (local extractive algorithm, zero API calls) when count exceeds threshold
• Prometheus + Grafana — p50/p99 latency dashboards, embedding API call duration, pool saturation
• MCP-compatible endpoint — Streamable HTTP server on port 8001; 4 tools (store, search, delete, count); connects to Claude Desktop and Claude Code
• HNSW index option — flag to switch from ivfflat to HNSW for deployments with >1M vectors
• Importance-weighted retrieval — factor importance score into ranking formula alongside similarity and recency

Tech stack