Your AI Coding Agent Forgets Everything. I Fixed It With a Local MCP Memory Server

Posted on Oct 6 • Originally published at github.com

Every new chat window, you re-explain your own codebase.

"We use PostgreSQL." "No, the auth middleware is in src/lib/auth, not src/api." "Yes, we already decided against Redux — that was three sessions ago."

It is the single biggest tax on AI-assisted development, and it has a boring name: agents have no memory. Context windows grew, but sessions still start from zero.

So I built Awareness Local — a local-first MCP memory server that gives your coding agent persistent memory across sessions. No cloud account. Works offline. Your memory is stored as plain Markdown.

npx @awareness.market/setup Enter fullscreen mode Exit fullscreen mode The installer detects your IDE (13+ supported: Cursor, Claude Code, Windsurf, Cline, Copilot, Codex CLI, Zed, JetBrains Junie, ChatGPT Desktop, …), wires up the MCP server, and starts a local daemon on 127.0.0.1:37800.

From then on your agent has five tools available:

Memory lives in.awareness/, as Markdown you can read, edit, diff, and commit:

.awareness/ ├── memories/2026-03-22_decided-to-use-postgresql.md ├── knowledge/decisions/postgresql-over-mysql.md ├── tasks/open/implement-rate-limiting.md └── index.db # SQLite FTS5 index, rebuilt automatically Enter fullscreen mode Exit fullscreen mode This is the whole point of local-first. Your agent's memory is not locked in a vendor's vector store. It is in your repo, in your git history, subject to your review. If the daemon dies, you have lost nothing — you have a folder of Markdown.

Retrieval is hybrid: SQLite FTS5 keyword search + local embeddings, fused with Reciprocal Rank Fusion. No LLM calls on the retrieval path, so recall costs you nothing per query and stays fully offline.

I evaluated it on LongMemEval (ICLR 2025) — 500 human-curated questions, ~115k tokens of history per question:

Recall@1 80.2% Recall@3 92.8% Recall@5 96.0% ◀ primary Recall@10 98.6% Enter fullscreen mode Exit fullscreen mode Method: Hybrid RRF (BM25 + vector), multilingual-e5-small, zero LLM calls on the retrieval path, run on an Apple M1 / 8GB RAM in ~35 minutes total.

The ablation is the interesting part — neither method alone gets there:

Vector-only: 92.6% BM25-only: 91.4% Hybrid RRF: 95.6% ← +3% over either single method Enter fullscreen mode Exit fullscreen mode Retrieval code and scripts are in the repo, so you can re-run it yourself rather than trust a table in a blog post.

The naive version of "give the agent memory" is to dump everything into context and watch your token budget evaporate. Awareness does a two-phase recall instead:

Phase 1 awareness_recall(query, detail="summary") → ~80 tokens per hit: title + summary + score → the agent picks what is actually relevant Phase 2 awareness_recall(detail="full", ids=) → full content, only for the selected items Enter fullscreen mode Exit fullscreen mode The agent decides what to load. You pay for relevance, not for volume.

Before: every session re-opens the same three questions.

After: the agent opens the session already knowing that you migrated MySQL → PostgreSQL, that you chose it for JSON support, and that two TODOs were left open.

That difference compounds. Long migrations, multi-week refactors, and team handoffs are where it stops being a convenience and starts being the reason the work is tractable at all.

npx @awareness.market/setup Enter fullscreen mode Exit fullscreen mode Repo (MIT): https://github.com/everest-an/Awareness-Market Docs: https://awareness.market/docs Discord: https://discord.com/invite/BMjkCKeqnJ Telegram: https://t.me/awarenessmarket Python and TypeScript SDKs ( wrap_openai() / wrap_anthropic() interceptors) and an OpenClaw plugin are in the same ecosystem if you want memory inside your application, not just your IDE.

If re-explaining your codebase to your agent is getting old, a star helps more developers find it: ⭐ https://github.com/everest-an/Awareness-Market