Skip to main content
Glama
ShmuelOps

hindsight-shared-memory

by ShmuelOps

hindsight-shared-memory

CI License: MIT

Always-on, local Hindsight memory for Claude Code on your Mac.

This repo runs the Hindsight server and its web UI as background daemons on localhost. Claude Code connects to them through the official Hindsight Coding Agents plugin. Every session remembers project decisions automatically, and every session and agent in a repo shares the same memory.

  • No API key. Fact extraction runs on your Claude Code login (Pro, Max, Team or Enterprise), using Haiku.

  • Automatic. The plugin injects relevant memories when a session starts and saves each session when it ends. You don't need to prompt it or keep CLAUDE.md rules.

  • Always on. launchd daemons for the server and the UI, started at login and restarted if they crash.

  • Local and private. Everything binds to 127.0.0.1 (upstream defaults to 0.0.0.0). It uses an embedded Postgres and local ONNX embeddings and reranking, and starts offline once the models are downloaded.

  • Light. Uses Hindsight's slim build, with no PyTorch (about 0.8 GB instead of 1.8 GB).

๐Ÿ‘‰ Install guide: from scratch on macOS in about 10 minutes.

How it works

 Claude Code (any repo) โ”€โ”€ hooks + MCP โ”€โ”€โ”   Hindsight Coding Agents plugin (npx)
 Codex / Cursor / โ€ฆ (same plugin) โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ค   recall on session start ยท retain on Stop
                                         โ–ผ
               hindsight-shared-memory daemon  โ”€โ”€ 127.0.0.1:8888  (launchd)
                 โ”œโ”€ LLM: claude-code provider โ†’ Haiku on your Claude login
                 โ”œโ”€ ONNX embeddings + FlashRank reranker (local, cached)
                 โ””โ”€ embedded Postgres (pg0) โ”€โ”€ 127.0.0.1:5488
               Hindsight Control Plane UI โ”€โ”€ 127.0.0.1:9999  (launchd)

The plugin gives each git repo its own bank (coding-agent::<repo>). The first session seeds it from git history and a short Haiku survey of the code. The bank then maintains knowledge pages such as the component map, conventions, key decisions and initiatives. It also adds hindsight_* MCP tools and a hindsight-coding-agent skill for explicit memory work.

Related MCP server: Unforget MCP Server

What this repo adds on top of Hindsight

Upstream default

Here

Bind address (API and UI)

0.0.0.0: your whole network, no auth

127.0.0.1

LLM

OpenAI (API key)

claude-code + claude-haiku-4-5, no key

Install size

1.8 GB (PyTorch)

~0.8 GB (ONNX + FlashRank)

Offline start

Checks Hugging Face on every start

Offline once the models are cached

Reranker cache

/tmp, wiped on reboot

~/.cache/hindsight-shared-memory

Embedded Postgres port

5432 (clashes with a local Postgres)

5488

Running it

You run it yourself

install-service / install-ui launchd daemons

Daemon PATH

โ€”

Stable dirs only, so a claude shim from a terminal session isn't baked in

Server MCP endpoint (/mcp/<bank>/)

~35 tools, including delete_bank

5 memory tools, for other MCP clients

CLI

hindsight-shared-memory serve               # run the server in the foreground
hindsight-shared-memory install-service     # server daemon (launchd; systemd --user on Linux)
hindsight-shared-memory uninstall-service   # remove it (memories are kept)
hindsight-shared-memory install-ui          # web UI daemon, pinned to the server version (needs Node.js)
hindsight-shared-memory uninstall-ui

Configuration

Every setting is a Hindsight environment variable. Set it before serve, or before install-service, which bakes your HINDSIGHT_* variables into the daemon.

Variable

Default here

Notes

HINDSIGHT_API_HOST

127.0.0.1

HINDSIGHT_API_PORT

8888

If you change it, pass the new --api-url to the plugin and the UI

HINDSIGHT_UI_PORT

9999

Used by install-ui

HINDSIGHT_API_LLM_PROVIDER

claude-code

Or ollama, openai, anthropic, โ€ฆ or none (storage and search only)

HINDSIGHT_API_LLM_MODEL

claude-haiku-4-5

Sonnet extracts more thoroughly but is much slower (~30 s per retain in testing)

HINDSIGHT_API_EMBEDDINGS_PROVIDER

onnx

Model intfloat/multilingual-e5-small

HINDSIGHT_API_RERANKER_PROVIDER

flashrank

Cached in ~/.cache/hindsight-shared-memory/flashrank

HINDSIGHT_API_DATABASE_URL

pg0://hindsight-shared:5488

Data in ~/.pg0/instances/hindsight-shared/. Or a postgresql:// URL

HINDSIGHT_API_MCP_ENABLED_TOOLS

retain,recall,reflect,list_memories,invalidate_memory

Only for the server's own /mcp endpoint; the plugin uses the REST API

The plugin's settings (bank routing, what to inject, survey options) are in ~/.hindsight/coding-agent.json; see the plugin docs. This setup changes one of them, autoInject: "recall", and the install guide explains why.

Security

  • The API and the UI have no authentication. They bind to 127.0.0.1 and must stay there, since anything that can reach the ports can read and write memories. CI checks that the server can't be reached on a non-loopback address.

  • Memories, and whole session transcripts sent by the plugin, are stored in plaintext in the local Postgres. Avoid pasting secrets into sessions.

  • The daemons keep only a stable PATH and HINDSIGHT_* variables from your environment.

See SECURITY.md.

Development

git clone https://github.com/ShmuelOps/hindsight-shared-memory && cd hindsight-shared-memory
uv sync
uv run pytest     # unit tests, plus a real server (LLM "none") checked for sharing and isolation over MCP
uv run ruff check . && uv run ruff format --check .

CI runs on Ubuntu and macOS. See CONTRIBUTING.md.

License

MIT. Hindsight is MIT-licensed by Vectorize.

Maintenance

ActivityMaintained
ResponsivenessNo issues

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    A
    maintenance
    Provides persistent memory for AI coding agents via MCP, enabling agents to store and semantically recall facts, events, and lessons across sessions, all running locally without cloud dependencies.
    Apache 2.0
  • A
    license
    Not graded
    quality
    C
    maintenance
    Long-term memory for Claude Code, Cursor, and MCP clients with zero-config embedded PostgreSQL. Stores and retrieves facts, preferences, and decisions using 4-channel semantic, BM25, entity, and temporal search with cross-encoder reranking.
    1
    Apache 2.0
  • A
    license
    Not graded
    quality
    B
    maintenance
    Provides a self-hosted shared memory service that lets AI agents capture and recall durable facts, decisions, and context across multiple tools and MCP-capable clients.
    3
    MIT
  • A
    license
    Not graded
    quality
    B
    maintenance
    Enables AI agents to store and recall durable facts, decisions, and lessons over MCP using hybrid semantic and keyword search, typed memories, graph relations, sessions, and time-travel queries. Runs locally with no server process or API key, persisting everything to files that both agents and a CLI can share.
    14 npm
    MIT