Cortex

Transparent, self-pruning memory for AI agents — the memory you can see, audit, and that doesn't bloat. MCP-native.

AI agents forget everything between sessions. The usual fix — dump all chat history into a vector DB and stuff it into the prompt — doesn't scale, never forgets noise, and isn't reusable. Cortex is different:

1. Decides what to remember — an LLM extracts only durable facts/preferences/decisions from each exchange (ignores small talk).

2. Lets the unimportant fade — every memory has a salience-scaled forgetting curve; throwaway details decay and get dropped, important facts resist forgetting.

3. Recalls what matters — retrieval ranks by similarity × strength, returning only the few memories that matter for the current request.

4. Is reusable infrastructure — exposed as an MCP server, so the same memory plugs into any MCP-capable agent (chat assistant, coding assistant, …).

Built for the Qwen Cloud Global AI Hackathon 2026 · MemoryAgent track.

How it's different

Agent memory is a crowded space (Mem0, Zep, Letta/MemGPT). Cortex's wedge isn't "we store memories" — it's the parts they under-serve:

• Active forgetting + consolidation — restating a fact merges (no bloat), recall reinforces it, contradictions supersede the old fact, and stale memories decay out. Not a fixed decay curve — adaptive. Proven by eval.py (89% footprint shrink, 0 important lost) and test_dedupe.py.

• Transparency & governance — memory you can see, audit, edit, and delete. Every memory carries a persisted, exportable audit trail (provenance + who accessed it + when + deletions), so you can answer "what does the AI know about this user, and what happened to it." The open gap for enterprise/regulated + consumer-trust use. Proven by test_audit.py.

• MCP-native — a drop-in memory primitive any agent inherits, no SDK lock-in.

Positioning: transparent, self-pruning memory — the agent memory you can audit, and that doesn't bloat.

Related MCP server: aimemory

Try it in 30 seconds (no API key)

Visual demo (browser):

open web/index.html

Hit Take the tour — watch memories form, get recalled, and (after simulated weeks) the throwaway ones fade and forget, then switch agents and see the memory carry over.

Prove the engine (terminal, no deps):

python3 demo_offline.py        # remember → retrieve → forget → recall-after-noise
python3 test_dedupe.py         # restatements MERGE (no bloat); recall reinforces
python3 test_shared_store.py   # a second, independent agent inherits the memory
python3 test_namespace.py      # per-user isolation + cross-agent sharing
python3 test_audit.py          # exportable per-memory audit trail + logged deletion
python3 eval.py                # keeps signal, forgets noise, 89% context shrink

These run offline using a deterministic mock embedding.

Run the API

pip install -r requirements.txt
cp .env.example .env            # set DASHSCOPE_API_KEY for live Qwen (optional)
uvicorn main:app --host 0.0.0.0 --port 8000
python3 -m cortex.mcp_server    # (separately) Cortex as an MCP server
python3 test_api.py             # HTTP smoke test (offline, mock embeddings)

Multi-tenant REST API — memory is isolated per namespace (a user id), shared across agents within it:

• POST /remember {namespace, content, kind, salience} · POST /recall {namespace, query}

• POST /chat (live Qwen) · GET /memories?namespace= · GET /audit?namespace= · DELETE /memories/{id}

• GET /health

With a key set, cortex.qwen_client uses Qwen models + real embeddings automatically; without one it falls back to the offline mock (so engine ops + the whole API work key-free for dev).

Architecture

Data flow — both the MCP tools and the REST endpoints go through the same registry -> MemoryEngine -> qwen_client/store core, so every caller (an MCP agent or a plain HTTP client) gets identical dedupe/decay/contradiction-resolution behavior:

flowchart LR
    subgraph Clients
        MCP["MCP client\n(remember / recall tools)"]
        REST["REST client\nPOST /remember · /recall · /chat\nGET /memories · /audit"]
    end

    MCP -->|"mcp_server.py"| REG
    REST -->|"main.py"| REG

    REG["registry.py\nper-namespace engine lookup\n(multi-tenant layer)"]

    REG --> ENG

    subgraph ENG["MemoryEngine (memory_engine.py)"]
        direction TB
        EXTRACT["extract / classify\n(consider(): new · duplicate · update)"]
        SCORE["score & decay\n(salience x recency half-life)"]
        DEDUP["dedupe / merge\n(cosine similarity)"]
        RETRIEVE["retrieve\n(similarity x strength, reinforce)"]
        EXTRACT --> DEDUP --> SCORE --> RETRIEVE
    end

    ENG -->|"embed() / chat()"| QWEN["qwen_client.py\nDashScope embed + chat\n(env-key-gated; mock fallback offline)"]
    ENG -->|"load() / save()"| STORE["store.py\nJSON persistence\n(memories + audit log)"]

cortex/
  memory_engine.py   dedupe/merge, salience-scaled forgetting, reinforce, supersede, audit
  store.py           persistent store (memories + audit log)
  registry.py        per-namespace engine registry — the multi-tenant layer
  config.py          env-driven configuration (12-factor)
  mcp_server.py      MCP server: remember() / recall()  ← the reusability layer
  qwen_client.py     Qwen via DashScope (OpenAI-compatible) + offline mock
  models.py          Memory model
main.py              multi-tenant FastAPI REST API
web/index.html       live memory-graph UI (multi-agent, audit drawer, inspector)
Dockerfile           container for deploy (Alibaba Cloud)

Why it maps to the judging rubric

• Innovation (30%) — delivered as an MCP integration; the remember/forget policy is a custom skill (both named in the rubric).

• Technical Depth (30%) — standalone memory engine (extract → score → decay → rerank) + shared store, modular and reusable.

• Impact (25%) — universal pain, shipped as open-source MCP infrastructure (pip install and plug in).

• Presentation (15%) — the live memory-graph UI literally visualizes the key logic; the cross-agent swap proves the claim.

Deploy

docker build -t cortex .
docker run -p 8000:8000 -e DASHSCOPE_API_KEY=... -v cortex-data:/data cortex

Target: Alibaba Cloud (a hackathon pass/fail requirement) — push the image and run it; mount a volume for /data.

Honest prod notes: the JSON-file store is fine for the hackathon/MVP but should be swapped for a managed store (Alibaba Tablestore / a vector DB) at real scale; the API is currently unauthenticated (add an API key / JWT layer before exposing it publicly). Both are clean swaps behind store.py / a FastAPI dependency.