Vibeleaderboard
Index / app
Visit github.com
Category
AI Agents
Rank

Previous survey · No. 784 ·

Pricing
Open Source
Type
APP
Use case
Data, Retrieval & Knowledge · Coding
Interfaces
Agent Skill / Plugin
Latest release
v3.4.7
Date

About

Long-term memory for coding agents, so you can stop Claude Code mid-task, start Codex in the same directory and carry on without re-explaining the architecture or the approaches that already failed. It keeps a git-versioned markdown wiki alongside SQLite indexes, retrieving with full-text search, entity matching, graph neighbours and optional vector similarity, and ranks decisions and procedures above ephemeral session notes. Lifecycle hooks capture prompts and tool calls across Claude Code, Codex, Kimi Code and Kiro CLI.

What it does

A self-hosted server that gives several coding assistants one shared, persistent notebook per project. As you work, hooks in each assistant quietly record what happened, scrub it for private data, and turn each finished session into a readable page plus a note for whoever picks up next. The next assistant, on any machine, gets a short briefing and can search everything. The plain text pages are the real record; the database is an index that can be rebuilt from them. Language model calls are optional, so capture, search and handoffs work with no API key.

Why it's ranked here

The breadth and the engineering discipline are both unusual. The project lists first-party integrations for more than twenty assistants and editors, ships packages for Arch, Fedora, macOS, Windows and Docker, and is MIT licensed. Its security document maps each isolation guard to a test that tries to break it, and most rows are rated strong. It also publishes retrieval scores from an in-repo harness on a public memory benchmark: 0.815 hit-at-5 with local embeddings and 0.666 with keyword search alone, with the method to reproduce them.

What's good

Handoffs between sessions are claimed exactly once, and tests check that a second assistant or another user cannot steal one. Memory you actually open or search fades more slowly, while compaction, deduplication and per-tier decay stay off until you enable them. Concurrent writes to the same page keep the older version instead of overwriting it, and a single page can be restored from a git checkpoint. Local embeddings run inside the binary with no egress. Destructive commands state precisely what they remove, and several refuse to run without an explicit confirm flag.

Tradeoffs

This is infrastructure, not a plugin: you run and maintain a server, and team use means managing users and tokens. The surface is large, with dozens of subcommands and many configuration switches. Support varies by assistant: several get only search tools with no automatic capture, and some lack a true session-end signal, so you must finalize sessions by hand. Native Windows is marked experimental. Capture trims each excerpt at 2 KB, so detail deep inside one long turn cannot be found. Pruning old raw observations is irreversible and blocks later re-summarizing.

How to use it well

Best for someone who switches between Claude Code, Codex and other terminal assistants on the same repositories, or a small team that wants shared project memory on its own hardware. Install the server, register it with each assistant, and add the capture hooks; on an older project, the bootstrap command can seed pages from history if you configure a model. Turn on local embeddings for noticeably better recall. Keep the wiki under git and take regular backups. It does not replace a general personal note vault; it remembers one repository at a time.

Technical notes+

Cargo.toml defines a Rust 2024-edition workspace (rust-version 1.95) of ten shipped crates plus test-support and an evals harness, with unsafe_code set to forbid. crates/ai-memory-store/src/lib.rs opens one SQLite file in WAL mode with a 5000 ms busy timeout, runs migrations with foreign keys off, then funnels every mutation through a single writer thread beside a four-connection read pool; new db directories and files are created 0700/0600 on Unix. docs/ARCHITECTURE.md describes memory_query fusing FTS5, entity match and link neighbours via reciprocal rank fusion, with optional vector cosine and a bounded authority multiplier favouring rules, decisions and procedures. docs/security-boundaries.md maps each guard to adversarial tests. crates/ai-memory-cli/src/cli.rs is a clap-derived CLI; crates/ai-memory-mcp/src/lib.rs exposes the MCP server.

Observed

License
MIT (LICENSE)
Language
Rust workspace, minimum toolchain 1.95 (Cargo.toml)
Interfaces
CLI binary and MCP server over stdio or HTTP, plus lifecycle hooks (crates/ai-memory-cli/src/cli.rs)
Storage
Git-versioned markdown wiki as source of truth; SQLite as rebuildable index (docs/ARCHITECTURE.md)
Platforms
Linux, macOS and WSL2 supported; native Windows experimental (docs/support-matrix.md)
Packaging
AUR packages, Fedora RPM, Docker images for amd64 and arm64, macOS and Windows release archives (docs/support-matrix.md)
Unsafe code
Forbidden workspace-wide by lint (Cargo.toml)
LLM requirement
None by default; LLM consolidation and reranking are opt-in (README.md)

Read from README.md, Cargo.toml, LICENSE, docs/ARCHITECTURE.md, docs/security-boundaries.md, crates/ai-memory-mcp/src/lib.rs, crates/ai-memory-store/src/lib.rs, crates/ai-memory-cli/src/cli.rs, docs/benchmarks/README.md, docs/support-matrix.md.

Intel on Ai Memory

More in Intel

Tech Stack

Rust

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.