Skip to content
coworkplugins
Comparison

Claude Code memory plugins compared: what each one keeps, and where

Ten Claude Code plugins that remember past sessions, shrink the context window or record decisions: what each stores, where it goes, what it needs.

By Arun BabuIntermediate

Claude Code starts each session fresh, and a long session loses detail when it compacts. A whole group of plugins tries to fix that. They don't all do the same thing, though: some remember past sessions, some keep the context window small, and some write decisions into your repository.

This page compares ten of them from this directory. Everything below comes from each plugin's own README and files, read on November 16, 2026. We have not installed or tested them, so treat this as a map of what each one says it does. Star counts are from GitHub on the same day.

The short version

Before installing any of them, read what to check at the end. Several send data off your machine or run code in the background on every turn.

Plugins that remember past sessions

Claude Mem (117,521 stars, Apache-2.0) records what the agent read, edited and ran as short notes, and searches them in later sessions. It uses hooks, MCP search tools, a search skill, a local background service and a second "observer" agent that writes the notes. Notes are stored locally in SQLite with vector search. Its README says cloud sync is opt-in but "anonymous usage analytics are on by default". The installer asks you to sign in through the browser, and the observer's notes are written by its own service, your OpenRouter or Gemini key, or your Anthropic plan. Install: npx claude-mem install, or add the thedotmack/claude-mem marketplace and install claude-mem. It also supports Codex, Cursor, OpenCode and several other agents.

Memex (7 stars, MIT) saves each session's work as searchable memos in an Obsidian-compatible vault on your machine and archives transcripts at each compaction. It uses hooks at session start, on each prompt, before and after compaction and at session end, plus one that scrubs secrets from tool output. An optional cloud copy runs on your own Cloudflare account, and semantic search needs a Gemini key or LM Studio. It needs Python 3.11+ and uv; install with uv tool install memex-plugin, then add the linxule/memex-plugin marketplace.

LORE (4 stars, AGPL-3.0, beta) keeps a capped set of facts about you and the project, beliefs backed by evidence, and a search over past transcripts, in a local SQLite store. Setup builds a Rust tool, so you need rustup, and it changes your Claude settings, including turning off Claude Code's built-in auto-memory, asking before each change. Its README notes that "derived beliefs are written without approval" and that its review step sends a scrubbed digest to your session's Anthropic endpoint. Install: add docwilde/lore, install lore, then run /lore:setup.

Obsidian Second Brain (5,344 stars, MIT) saves conversations as linked notes about people, projects, decisions and tasks in your Obsidian vault, and recalls them later. It is large: 49 commands, a skill, hooks, a vault MCP server and optional scheduled agents. The README says the core commands need no API keys and your notes stay as Markdown files on your computer; its research commands do call outside services such as Perplexity and Gemini. It needs git, uv and an OBSIDIAN_VAULT_PATH setting.

Plugins that keep the context window under control

Context Mode (34,165 stars) keeps large tool output out of the context window: it indexes the output locally and hands the agent only the part it asks for. It uses hooks, an MCP server and skills, and stores one SQLite file per project in ~/.context-mode. The README says "Zero telemetry" and lists the four network calls it makes, each started by you or the agent. Two things to know: code run through its ctx_execute tool "runs on your host as your user", and its licence is ELv2, which is source-available rather than open source. It needs Node 24.15 or later.

The Context Kit (1 star, MIT) writes the session's reasoning to .claude/session-ledger.md before compaction, puts it back at the next session start, and refuses a compaction from an unprepared state. It uses hooks plus /checkpoint, /brief and /prove commands, and the README says an install "is offline and standalone". It needs bash and python3, doesn't run on Windows outside WSL, and its trigger phrases are English and German only.

fast-jev-compaction (7,980 stars, MIT) replaces Claude Code's compaction summary: it removes stale tool calls and results and keeps the rest word for word. The catch is in its README: "The state sent to Jev is the whole conversation so far". The Jev model runs on a third-party service, so you need a TYPESAFE_API_KEY, and it relies on an early-access Claude Code setting (CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1, v2.1.274 or later).

Plugins that keep decisions with the code

kgai (7 stars, MIT) records a team's structural code decisions, including what was rejected and why, in a knowledge graph stored in the project's .kgai folder. The README says it is "local-first", with optional team sync to your own S3. It downloads an engine binary on first use, may add a line to your shell profile, and after any turn that edits code its Stop hook continues the turn to capture the decision. Linux and macOS only, or Windows through WSL.

Cadence (11 stars, MIT) is more a planning workflow than a memory plugin: requirements, decisions, plans and results live in your repository under .planning/, so a new session picks up where the last one stopped. It ships many agents and about 37 skills. The README says it "adds no telemetry". Note that you install it from a different repository's marketplace, crenshawdev/foundation.

A plugin that learns skills

AutoHarness (11,134 stars, MIT) watches your real sessions and writes Claude Code skills from them, then merges, updates and prunes the skills it wrote. It uses hooks, an MCP server and two agents that run on a small model. It writes into .claude/skills/ in the background, and uninstalling leaves its state and those skills on disk. It needs Python 3.11 or later as python3; its README warns that the 3.9 version Xcode installs turns every hook off.

What to check before you install one

Memory plugins see more of your work than most, because remembering is their job. From the READMEs above, the questions that separate them:

  • Does anything leave your machine? Claude Mem has analytics on by default. fast-jev-compaction sends the conversation to a third-party API. LORE sends a scrubbed digest to your Anthropic endpoint. Context Mode, The Context Kit, Cadence and kgai say they keep data local.
  • Does it need a key or an account? fast-jev-compaction needs a paid API key, and Claude Mem's installer asks you to sign in. Memex's semantic search needs a Gemini key or a local model.
  • What runs on every turn? Most of these use hooks that run at session start, on each prompt or at compaction. kgai's continues a turn after code edits, and AutoHarness writes skills in the background.
  • What does it change? LORE edits your Claude settings, kgai may edit your shell profile, and AutoHarness writes into .claude/skills/.
  • What is the licence? Most are MIT or Apache-2.0. Context Mode is source-available (ELv2) and LORE is AGPL-3.0. Check if that matters to your team.

Each plugin's page in this directory links to its repository. Read the README and its hooks before installing, and start in a test project. How to install from a marketplace, and how to remove one cleanly, is in how to install Claude Code plugins and turn off or remove a plugin. What plugins can access in general is in plugin permissions.

Sources

Read more