iai-personal-memory-engine

English | 中文

---

What it is

A local server that speaks the MCP protocol and gives Claude, and any other MCP-compatible agent, 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 and turn. You never have to say "remember this" or "what did we say last time?".

I built this for myself. It worked. The benchmarks were mostly for my own curiosity.

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.

And unlike cloud memory services, there's no API key, no account, and no telemetry: the engine, the store, and the embeddings all run locally. The only things that leave your machine are the normal model calls your CLI already makes, a nightly consolidation step that asks your model for a single insight through the same subscription your CLI already uses (on by default; IAI_MCP_REM_DISABLED=1 turns it off), a daily version check, and the one-time download of the embedding model.

---

Quick start

Claude Code

python3.12 -m pip install -U iai-pme

Then run inside Claude Code:

/plugin marketplace add CodeAbra/iai-personal-memory-engine
/plugin install iai-memory@iai-pme

Restart the session, then verify:

iai --version
iai-mcp daemon status
iai-mcp doctor

Python 3.11 is also supported.

macOS or Linux: all-in-one source install

curl -fsSL https://raw.githubusercontent.com/CodeAbra/iai-personal-memory-engine/main/scripts/bootstrap.sh | bash

This builds the Rust engine and TypeScript wrapper, installs the background service and hooks, registers Claude Code, and runs the health check. It requires Git, Python 3.11/3.12, Node.js 18+, and Rust. To inspect the steps without changing anything:

curl -fsSL https://raw.githubusercontent.com/CodeAbra/iai-personal-memory-engine/main/scripts/bootstrap.sh | bash -s -- --dry-run

Other hosts

python3.12 -m pip install -U iai-pme
iai-mcp crypto init
iai-mcp daemon install
iai-mcp capture-hooks install --target codex

Replace codex with cursor, antigravity, hermes, openclaw, or all. MCP tools work with any MCP-over-stdio client; automatic capture and context injection depend on the hooks exposed by the host. See the technical reference.

New stores use the native engine format by default; an existing store keeps its current format on upgrade. To move an existing legacy SQLite store onto the native engine, run iai-mcp migrate-to-lilli — iai-mcp doctor prints the exact command, and the technical reference documents the full flow.

---

What happens after installation

Event	Action
Prompt	New turns are appended to a session buffer as file IO; no embedding or engine RPC is needed on the capture path
Session end	Remaining transcript content is rolled over for ingestion; hook failures do not block the host
Session start	A bounded memory prefix is exposed as host context; an empty store or unavailable engine yields empty output
Later turns	Supported hosts receive a small foresight or delta pack with age and revision markers
Idle time	Captures are embedded, deduplicated, encrypted, inserted, clustered, consolidated, reinforced, and decayed

The background process is called the daemon in the CLI. The MCP wrapper and iai can still read the local store directly when it is asleep or temporarily unavailable.

---

How it works

Memory model

Tier	Contains
Episodic	Timestamped, write-once fragments of what was said
Semantic	Summaries induced from related episodes during idle consolidation
Procedural	Ten bounded behavioural parameters learned over time

Distinct hyperdimensional representations keep literal detail, semantic structure, and behavioural tendencies from collapsing into one vector surface.

The local, LLM-free recall path combines semantic similarity, graph evidence, recency, temporal validity, and lexical evidence. memory_recall returns both hits and anti_hits; memory_contradict closes the old record's validity interval, creates a new record, and links the two.

While idle, the engine groups related episodes, induces semantic memory, reinforces useful paths, and decays weak unreviewed edges. One REM step, on by default, may invoke claude -p through the user's existing Claude subscription, capped at no more than 1% of the daily quota; IAI_MCP_REM_DISABLED=1 turns it off. No Anthropic API key is required.

First-party components

Component	Role
Hippo	Encrypted records, vector index, and graph in one local store
MOSAIC	Leiden-family community detection with stable community identity
Lilli HD	Hyperdimensional substrate and structural recall
Native engine	Rust embedder and graph kernels

---

Dashboard and CLI

iai brain

The local dashboard searches the store, exposes graph neighbourhoods and contradictions, pins or fades memories, ingests files, controls the background engine, and reports token-use estimates from your own store.

iai recall · temporal-recall · search · ask · capture · teach · upload
iai watch · brain · status · last

iai upload accepts documents, Office files, e-books, source code, configuration files, and directories. Full formats and administrative commands are listed in docs/REFERENCE.md.

---

Benchmarks

Every harness ships in bench/; methodology and reproduce commands are in BENCHMARKS.md.

Benchmark	Result
Rescue@10 after contradiction	1.000
Historical-verbatim hit@10	1.000
LongMemEval-S R@5, product embedder	0.962
LongMemEval-S R@10, product embedder	0.978

Historical-verbatim retrieval uses a flat-cosine baseline of about 0.71. With the matched all-MiniLM-L6-v2 embedder, iai-memory and mempalace v3.3.6 both score R@5 0.966 and R@10 0.978; no win is claimed.

On the author's store, an automatically injected memory pack averaged about 350 tokens versus about 2,850 tokens for the agent-search round trip it replaced: approximately 88% cheaper on that measured workload. This does not apply to explicit memory_recall, whose default response budget is 1,500 tokens.

---

MCP tools

memory_recall              memory_temporal_recall
memory_recall_structural   memory_search
memory_capture             memory_contradict
memory_reinforce           memory_consolidate
profile_get_set            topology
schema_list                events_query
episodes_recent            curiosity_pending
claim_check

Fifteen tools cover cue, temporal, structural, and lexical recall; checking a claim against memory; capture and correction; reinforcement and consolidation; behavioural-profile control; and store introspection.

---

Compatibility

Host	Ambient behaviour
Claude Code	Session-start recall, per-turn updates, turn capture, and session capture
Codex CLI	Full integration through Codex hooks
Cursor	Session-start recall and capture; no per-turn text injection
Antigravity	Recall per invocation and lossless transcript capture
Hermes 0.5.0+	Recall before model calls and capture from its message store
OpenClaw	MCP tools on request; no ambient shell hooks
Gemini CLI and other MCP hosts	MCP tools; no bundled host-specific hooks unless listed above
Claude Desktop	MCP tools; plain Chat does not expose Claude Code-style ambient hooks

---

Privacy and limitations

• Records are encrypted at rest with AES-256-GCM. The store and key live under ~/.iai-mcp/; back them up together. Two kinds of files there are plain text: the per-turn capture spool until the engine's next pass encrypts it, and the caches the hooks inject memory from. The plugin README lists them.
• macOS and Linux use a Unix socket. Windows uses an ephemeral loopback port with a per-user token.
• There is no iai-memory account, telemetry pipeline, hosted dashboard, or cross-machine sync.
• iai-memory goes online in three places, each with an off switch: the nightly REM claude -p step (on by default, IAI_MCP_REM_DISABLED=1), a daily PyPI version check (IAI_MCP_VERSION_CHECK=0), and the first-use download of the embedding model from Hugging Face (IAI_MCP_EMBED_OFFLINE=1).
• The engine checks that your Claude Code login is a subscription by reading Claude Code's login record (~/.claude/.credentials.json, or the macOS Keychain item) when the MCP server starts and before the REM step. Nothing from the record is sent anywhere.
• The MCP server starts the engine in the background, and it keeps running after the session ends. Stop it with iai-mcp daemon stop.
• The store refuses to mix incompatible embedding generations; changing the embedder requires an explicit migration.
• Recall is usually mediocre during roughly the first ten sessions, and quality and latency depend on corpus size, language, embedder, and stored history.
• The default store is English-first. Raw non-English records require an explicit raw:<lang> tag and a multilingual or custom embedder.
• Windows support is beta. Ambient behaviour varies with host hook support.
• The project is solo-maintained and has no enterprise SLA.

Health and updates:

iai-mcp doctor          # 38 checks
iai-mcp daemon status
iai-mcp self-update

---

About the name

IAI — Independent Autistic Intelligence describes the memory design.

• Independent: the engine, store, embeddings, and dashboard run locally.
• Autistic: literal preservation, precise cues, sustained focus, and rare events retained as rare rather than smoothed into a typical summary. This is an operational design description, not a diagnosis or casual metaphor.
• Intelligence: used in the systems sense — a process that observes, adapts, reorganizes itself, and remains viable over time.

“Personal memory engine” describes the scope: one person's memory, on one machine, used by the assistant they already have.

---

Documentation

• docs/REFERENCE.md — technical and operational reference
• BENCHMARKS.md — methodology and reproduce commands
• docs/EMBEDDERS.md — providers, languages, and migrations
• CHANGELOG.md — release history
• CONTRIBUTING.md — development and test setup
• SECURITY.md — private vulnerability reporting

Issues and pull requests are welcome. Changes to retrieval, capture, contradiction handling, or consolidation should include relevant benchmark reruns.

Authors

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

License

MIT