iai-pme

The best open-source personal memory engine for AI coding assistants.

Every claim ships with the harness that proves it. Run the benchmarks yourself.

iai-pme

Your AI assistant forgets you every session. iai-pme gives it a memory that doesn't.

Independent Autistic Intelligence — a personal memory engine. Fully local, ambient. Works with Claude and other MCP-compatible hosts.

Table of contents

• What it is
• Quick start
• Usage
• How it works
• Built our own
• Benchmarks
• Configuration
• Doctor
• Notes for AI assistants
• Status and limitations
• Compatibility
• About the name
• Authors
• License
• Contributing

What it is

A local server that speaks the MCP protocol and gives Claude, and any other MCP-compatible assistant, a long-term memory. It captures every turn of every session verbatim, organizes those captures over time into a personal map of who you are, and serves a small slice of relevant memory back at the start of each new conversation. You never have to say "remember this" or "what did we say last time?".

I built this for myself. It worked. I've been running it daily for months, and now I'm sharing it. The benchmarks were mostly for my own curiosity. I wanted to know if it actually works or if I'd just gotten used to it.

Under the hood it's not a wrapper around someone else's vector store and graph library — the parts that matter are my own code: the storage engine, the community-detection algorithm, the hyperdimensional memory substrate, and a native engine that makes it fast. More on that in Built our own.

And unlike cloud memory services, nothing leaves your machine: no API key, no account, no telemetry. The local engine, the store, and the embeddings are all local.

Pick your path

Quick start

Prerequisites

• macOS (Apple Silicon tested)
• Python 3.11 or 3.12
• Node.js 18+
• A Rust toolchain — the native engine builds from source
• An MCP-compatible CLI host — Claude Code, Codex CLI, Gemini CLI, Cursor CLI, and others
• ~500 MB free disk

Windows and Linux aren't supported yet — the engine and its native core are macOS-only for now. Contributions are very welcome: if you'd like to port iai-mcp to Linux or Windows, open an issue or PR and I'll help however I can.

Install

git clone https://github.com/CodeAbra/iai-personal-memory-engine.git
cd iai-personal-memory-engine
python3.12 -m venv .venv && source .venv/bin/activate
pip install .

pip install builds the native Rust engine (iai_mcp_native — the embedder + graph kernels) automatically, as part of the package build, via setuptools-rust. There's no separate build script. If you change the Rust source later and need to rebuild it by hand, there's an escape hatch:

iai-mcp build-native        # rebuild the native engine in place

Then build the MCP wrapper and set up the local engine (it runs in the background):

cd mcp-wrapper && npm install && npm run build && cd ..
iai-mcp daemon install      # launchd on macOS, systemd on Linux
iai --version

Install the capture + recall hooks

This is what makes memory ambient. Without these hooks iai-mcp reads memory but never writes conversation content and never injects recall at session start. One command wires all three:

iai-mcp capture-hooks install       # copies all three hooks + patches ~/.claude/settings.json
iai-mcp capture-hooks status        # verify: should print "status: ACTIVE"
iai-mcp capture-hooks uninstall     # clean removal if ever needed

For Codex:

iai-mcp capture-hooks install --target codex

To install both:

iai-mcp capture-hooks install --target all

What the install does:

• Copies three hook scripts from deploy/hooks/ to ~/.claude/hooks/ (chmod +x):
  • iai-mcp-turn-capture.sh (UserPromptSubmit, timeout 5s) — appends each prompt + the preceding assistant turn(s) to a per-session buffer as pure file IO. Zero engine RPC during the session.
  • iai-mcp-session-capture.sh (Stop, timeout 35s) — at session end, rolls the buffer over for the local engine to drain, and runs iai-mcp capture-transcript --no-spawn as a safety net.
  • iai-mcp-session-recall.sh (SessionStart, timeout 30s) — calls iai-mcp session-start and pipes the assembled memory prefix to stdout, which Claude Code injects as additionalContext before the first prompt. Fail-safe: empty store or unreachable local engine yields empty stdout — session start is never blocked.
• Registers iai-mcp in Claude Desktop's config if installed.
• Idempotent — re-running detects existing entries and makes no changes.
• No secrets, no tokens, no network calls.

What happens at runtime:

• Every prompt (per-turn hook): appends new transcript turns to the session buffer. ~5 ms per turn, no embedding, no engine socket.
• Every session end (Stop hook): rolls the buffer over, captures any remaining turns. Fail-safe exit 0.
• Every session start (recall hook): assembles the cached memory prefix and pipes it to Claude. Empty store or unreachable local engine → empty stdout.
• When idle (local engine): drains the buffer through the shield → embed → dedup → encrypted insert pipeline on the WAKE → DROWSY edge (5-min idle) and after every REM cycle.

Connect your MCP host

Claude Code:

claude mcp add iai-mcp -- node "$(pwd)/mcp-wrapper/dist/index.js"

Or edit ~/.claude.json directly:

{
  "mcpServers": {
    "iai-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/iai-mcp/mcp-wrapper/dist/index.js"]
    }
  }
}

Use the absolute path. ~ and $HOME won't expand here.

For Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json.

Codex CLI:

[mcp_servers.iai-mcp]
command = "node"
args = ["/absolute/path/to/iai-mcp/mcp-wrapper/dist/index.js"]

[mcp_servers.iai-mcp.env]
IAI_MCP_PYTHON = "/absolute/path/to/iai-mcp/.venv/bin/python"
IAI_MCP_STORE = "/Users/you/.iai-mcp"

Codex hooks are stable in current Codex CLI builds. If hooks are disabled bylocal policy or an older install, enable [features].hooks = true in~/.codex/config.toml.

Verify

iai-mcp doctor
iai-mcp daemon status

Restart Claude Code. Start a session, do some work, exit. Then:

tail ~/.iai-mcp/logs/capture-$(date -u +%Y-%m-%d).log

You should see a rc=0 line. That's your first memory.

Usage

You do not call iai-mcp directly during a session. Once it's connected:

Capture is automatic. Every turn, yours and the assistant's, is recorded verbatim with timestamps and session metadata. You don't say "remember this."

Recall is automatic. When a new session starts, the local engine assembles a small relevant slice of your history and injects it into the conversation prefix. You don't say "what did we say."

Consolidation runs idle. Between sessions, the local engine merges duplicates, strengthens recall pathways for things retrieved often, and prunes weak edges. The system gets quietly better at remembering you over time.

After a few weeks of regular use the difference becomes noticeable. The assistant stops asking the same orientation questions, references things you mentioned in passing, and adapts to your style without being told.

There's also a CLI — you don't need it for normal use, but when you want to query or add to your memory straight from the terminal, iai is there: recall, capture, ask (LLM synthesis grounded in your memory), status, and last.

How it works

The local engine is a Python process that runs in the background — it sleeps when idle and wakes when your assistant needs it, so it isn't always-on or constantly using CPU. Your MCP client connects to it via a Unix socket. No network exposure.

Recall doesn't depend on the engine being awake. The store itself is always available: when the engine is asleep or not running, your assistant (and the iai CLI) read memory directly from the local store. The engine handles the fast LLM-free recall path when it's up, plus the nightly consolidation pass — it's never a gatekeeper on your memory.

Memory is stored in three tiers:

Episodic is verbatim, timestamped fragments of what was said. Write-once, never overwritten or rewritten.

Semantic is summaries induced from clusters of related episodes during idle-time consolidation.

Procedural is a small set of stable parameters about you, learned over time: preferences, style cues, recurring patterns. Eleven sealed knobs that shift based on what works.

The three tiers are backed by a hyperdimensional memory substrate — each kind of memory gets its own representation, so episodic detail, semantic gist, and procedural patterns don't collapse into one undifferentiated blob.

A background pass runs periodically (sleep cycles): it clusters episodes with my own community-detection algorithm, builds semantic summaries, decays old unreinforced connections, and reinforces frequently co-retrieved paths. Things you haven't revisited fade naturally. One step per night can make a single LLM call through your existing Claude subscription (claude -p) — no separate API key, capped at ≤1% of your daily quota. (iai-mcp doctor row (p) verifies there's no API-key SDK path installed at all.)

Recall combines three signals: semantic similarity, graph-link strength, and recency. All ranked together. The hot path runs entirely locally with no LLM in the loop.

All records are encrypted at rest with AES-256-GCM. The key lives in ~/.iai-mcp/.key (mode 0600). Back it up. Lose the key, lose the memories.

Everything lives at ~/.iai-mcp/. Embeddings are computed locally. The only data that leaves the machine is your normal conversation with whatever LLM API your client uses.

Built our own

Most memory projects are a thin layer over an off-the-shelf vector store and someone else's graph library. This one isn't. The load-bearing pieces are my own code, written for this exact workload — a small memory graph that mutates every night and gets queried on every recall:

These sit on a thin layer of proven, permissive primitives — SQLite, the candle tensor library, NumPy, and the audited cryptography AES implementation. I build the engine and the algorithms; I don't reinvent a database, a tensor kernel, or — deliberately — a crypto primitive. The interesting bricks are mine; the foundation under them is boring, battle-tested, and permissive. MIT throughout.

I wrote these because the off-the-shelf options were built for a different problem — large static graphs, multi-tenant clouds, gist-style summarization — and they were slower and a worse fit for "one person's memory on one machine, reorganized every night." Mine are faster on this shape of problem, which is the only shape I care about.

Benchmarks

I built these because I wanted honest numbers, not a leaderboard. Every harness ships in bench/ with a one-line reproduce command — run them and get your own results. Where a number missed its target or regressed, it says so. Full detail in BENCHMARKS.md.

LongMemEval-S — the one head-to-head arena

Validated in a single harness against mempalace on the identical 500 cleaned questions, session granularity, recall_any@k, raw (no rerank):

On raw retrieval — the headline both projects ship — it's an exact tie on the matched embedder — R@5 0.966 = 0.966 and R@10 0.978 = 0.978. Our product embedder scores 0.962 R@5, a 2-question-in-500 difference (noise). No win claimed — an honest tie is the strong, defensible statement. LongMemEval is a cold, one-shot retrieval test; it doesn't exercise cross-session memory, which is where the design's real edge is.

Where it actually leads — longitudinal memory

Cost & footprint (honest disclosure — not a brag)

One honest gap: retrieving the superseded wording of an updated fact verbatim regressed (0.90 → 0.71) in this release — separate from Rescue@10 (current-fact retrieval, still 1.000) — and is a tracked fix for the next release.

python -m bench.longmemeval_blind            # LongMemEval-S (raw)
python -m bench.contradiction_longitudinal   # Rescue@10 / longitudinal
python -m bench.personal_fact_drift          # drift / retention
python bench/sleep_ablation.py               # sleep-consolidation recall
python -m bench.tokens                       # session-start token cost
python -m bench.neural_map                   # recall latency
python -m bench.memory_footprint             # RAM footprint

Measured on an Apple M2 Max (64 GB). The harnesses are the proof — run them yourself.

Configuration

The old IAI_MCP_EMBED_MODEL knob is gone — the embedder is a single built-in English-only model. There are many internal tuning knobs (IAI_MCP_*), but you shouldn't need to touch them.

Doctor

iai-mcp doctor runs 23 checks against the local engine, the store, the native engine, and the runtime state. Output is one line per check: PASS, WARN, or FAIL.

iai-mcp doctor

What it checks:

A full-PASS run is healthy. Dropping (b) during a sleep cycle is normal (the socket is busy during consolidation). Multiple FAILs, or a FAIL on (a)/(f)/(v), means something is actually wrong.

Troubleshooting

Notes for AI assistants helping with installation

If you're an AI model helping a user install or debug iai-mcp, read this before diagnosing anything.

Terminology note: this doc calls the background component the local engine. Its process, CLI subcommands and doctor check labels still use the original name daemon — e.g. iai-mcp daemon status, pgrep -f iai_mcp.daemon, and the (a) daemon process alive check. "Local engine" and "daemon" refer to the same process.

The local engine has four FSM states: WAKE, TRANSITIONING, SLEEP, DREAMING. Separately, the wrapper-layer lifecycle has four states: WAKE, DROWSY, SLEEP, HIBERNATION. These are two different state machines in two different files. Don't confuse them.

Things that look broken but are not:

• last_tick_at hasn't changed in hours. This is a sleep-cycle tick timestamp, not a process heartbeat. The daemon can be alive and healthy with a stale last_tick_at. Check iai-mcp daemon status instead.
• fsm_state: DREAMING for hours. Normal. The local engine is running a sleep cycle. Wake transitions are event-driven (a new session arriving wakes it), not time-driven.
• status: deferred in the capture log. This is the only path the Stop hook takes. It writes captures to ~/.iai-mcp/.deferred-captures/ and the local engine picks them up on its next tick. This is by design, not a failure.
• iai-mcp doctor drops the socket check during SLEEP or DREAMING. The socket is intentionally unresponsive during consolidation passes. This is healthy.
• Cold start takes a few seconds. The local engine initializes the store and loads the embedding model into memory. Don't kill it because it hasn't responded in 5 seconds.

Actual problems look like: iai-mcp daemon status returning ok: False, doctor showing multiple FAIL results, or the daemon process not existing at all (pgrep -f iai_mcp.daemon returns nothing).

When in doubt, run iai-mcp doctor and read what it says. The output is self-explanatory.

Status and limitations

v1.0.0 — out of experimental. I built this for myself and ran it daily for months; it's now a stable release with a committed public surface. The MCP tool set and the on-disk store stay stable across 1.x — breaking changes go through the changelog with a deprecation window. It's still a solo-maintained project with no enterprise SLA, but it's no longer a moving target.

Limitations worth knowing about:

• English-only by design. The assistant translates to English on the way into memory; the store and the embedder are English-only on purpose.
• No cross-machine sync. The data lives where the local engine runs. Backup is cp -a ~/.iai-mcp/ somewhere safe.
• No GUI. Inspection happens through CLI subcommands (iai-mcp doctor, iai-mcp daemon status, iai-mcp topology).
• Cold start on a freshly booted machine takes a few seconds while the local engine initializes caches.
• Recall quality on the first ~10 sessions is mediocre. The system needs material to consolidate before it gets useful.

Compatibility

iai-mcp talks to its host over MCP-over-stdio — the same protocol every MCP-compatible CLI speaks. So the memory tools (recall, capture, ask, status) work with any MCP CLI:

• Claude Code — primary host, validated in daily use.
• Codex CLI — supported, with ambient capture through a Stop hook.
• Gemini CLI, Cursor CLI, and other MCP-over-stdio CLIs — connect through the same standard protocol; the MCP tools work out of the box.
• Claude Desktop — works; uses claude_desktop_config.json instead of ~/.claude.json.

Ambient capture (the hooks that record and recall automatically) ships for Claude Code and Codex today. On other CLIs the MCP tools work directly; wiring up their native hooks for fully automatic capture is a great first contribution — open an issue or PR.

About the name

The project is iai — a personal memory engine. The short name is an acronym; the descriptor says what it is.

IAI — Independent Autistic Intelligence (the memory style):

• Independent. Fully local. The local engine runs on your machine, embeddings are computed locally, no telemetry, no cloud dependency. Your memory is your data and stays your data — and it tunes itself over time without you steering it.
• Autistic. Describes the memory style, not a diagnosis or a metaphor. The memory is built around verbatim recall, attention to specific cues, and a refusal to smooth rare events into typical ones. Most memory systems compress and summarize aggressively, aiming to give the assistant a gist of the past. This one preserves what was actually said and surfaces it on a precise cue. In practice that shows up as: literal preservation over paraphrase; deep focus on the current thread rather than diffuse association; direct, unmasked output; a stable identity that doesn't drift. The trade-off is intentional: more storage and a stricter retrieval interface, in exchange for not losing details.
• Intelligence. Used in the systems sense — something that observes, adapts, and stays viable over time — not the marketing sense.

Personal memory engine (what it is): not a chatbot feature or a cloud add-on, but a memory engine — its own storage, clustering, hyperdimensional substrate and native core — that belongs to one person and runs on one machine. See Built our own.

It's an operational design choice about how memory should behave, not a clinical claim.

Authors

By Areg Aramovich Noya, in collaboration with the team at lcgc.dev.

I built this because I needed it. It works for me. If it works for you, take it.

License

MIT

Contributing

Issues and PRs welcome. If your change touches retrieval, capture, or consolidation, include bench re-runs.