A guide to memory add-ons for Claude Code

Which memory plugin should you use with Claude Code?

Updated 20 October 2026 · each system checked against its maker’s docs that day

Start with what Claude Code already has. CLAUDE.md holds the rules you write and auto memory holds the notes Claude keeps for itself, both as files on your machine. If you want every session written up for you without asking, add claude-mem. For memory of the codebase and the decisions in it, those two beat everything else here, ours included. Mem0, Supermemory and Cognee bring memory services of their own, through a cloud account or a server you run. Counterparts is for a different job: Claude’s own memory of you and the relationship, which fades unless it comes up again.

We make Counterparts, a memory for Claude, and it’s one of the entries below.

Side by side

The options side by side

Each name goes to its page in our directory, rated by the same rules as every other system.

Option	Remembers	Who writes it	Where it lives	Account or key	Cost
Built in: CLAUDE.md and auto memory	Your project rules; notes on you, your corrections and decisions	You write CLAUDE.md; Claude writes auto memory	Plain files on your machine, one folder per repo	Just your Claude Code login	Included
claude-mem	What happened in each coding session	A second AI that watches each step (Claude, on your plan, by default)	Your machine; synced to cmem.ai once you sign in	The installer asks you to sign in; you can skip it	Free; Pro $30/mo
Engram Alpha	A project’s decisions, problems and fixes	Your coding assistant; you can review and edit	One file in your repo	None	Free (MIT)
trusty-memory	Notes and facts the agent chooses to save	The agent, through its memory tools	A local server on your machine	None (a key is optional, for one extra step)	Free (MIT)
Mem0	Short facts from your sessions, per project and per person	Mem0’s extraction AI	Mem0’s cloud	A Mem0 API key	Free tier; from $19/mo
Supermemory	Your conversations and tool use, turned into facts	Its own models on the cloud; the model you bring, self-hosted	Supermemory’s cloud, or a server on your machine	A Supermemory API key	Free ($5 of credits a month); Pro $19/mo
Cognee	Prompts, tool steps and answers, as a knowledge graph	An AI that extracts facts (your key, or your Claude plan)	A local server by default, or Cognee Cloud	An AI model key, or your Claude plan	Free locally; Cloud $1 per million tokens after a free 10M
Counterparts	You, your work together, and itself; not the codebase	Claude, in its own words, in your session; a nightly run on your plan tidies them	One SQLite database on your machine	None	Free (MIT)

Prices are each maker’s own, read on 20 October 2026. “Free” means the tool; anything that runs an AI to write or tidy notes uses your Claude plan or a model key.

By what you need

Which memory plugin should you use?

If Claude should follow your project’s rules every time, write a CLAUDE.md (/init drafts one) and commit it. If you keep correcting the same things, leave auto memory on, or say “remember that …”. Neither needs a plugin; our guide to Claude Code’s built-in memory covers both.

For a record of every session kept for you, use claude-mem. To keep the project’s decisions and fixes as something you can open and edit, try Engram Alpha; for a local MCP server that holds notes the agent saves, trusty-memory.

If the same memory should reach other tools or an app you’re building, and a cloud account is fine, look at Mem0 or Supermemory. Cognee fits when you want a knowledge graph on your own machine.

If you want Claude to remember you, how you work together and who it has been with you, that’s what Counterparts is for. It won’t remember your codebase.

In detail

Each one, and what goes wrong

The problems below come from each project’s own docs and issue tracker, read on 20 October 2026. A closed issue is marked as closed.

Built in: CLAUDE.md and auto memory

Beside the CLAUDE.md files you write, auto memory is on by default: Claude saves notes on who you are and how you work, corrections you gave it, project decisions the code doesn’t show, and where to find things outside the project. They go in ~/.claude/projects/<project>/memory/, one folder per repo. For project rules and codebase decisions, nothing below beats it.

What goes wrong: Nothing fades. Requests to let unused notes decay (#59111) and about a wrong belief auto memory kept bringing back (#70525) were auto-closed for inactivity. Our guide to stale AI memory covers which systems do let unused notes fade. Only the first 200 lines or 25KB of the MEMORY.md index load each session; Claude Code now warns Claude when the index nears that limit and returns an error when it’s over. That warning reaches the model, not you, and a session can’t tell whether its index loaded whole (open, #82056). Notes stay on one machine; the request to change that has 45 reactions (#25739).
See what it stored: /memory lists your CLAUDE.md files and opens the auto memory folder; /context shows which CLAUDE.md files loaded.

Claude Code memory in the directory · Anthropic’s docs

claude-mem

A plugin that takes notes on your coding sessions without being asked. A second AI, the observer, watches each tool step and writes short notes, plus a summary after each reply. The next session opens with the newest; Claude searches the rest. They live in SQLite under ~/.claude-mem/. The plugin is free (Apache-2.0); CMEM Pro, $30 a month after a 30-day trial, runs the observer on their servers. With about 102,000 GitHub stars, it’s the most-starred plugin on this page.

What goes wrong: The installer asks you to sign in, and signing in turns on cloud sync, which uploads your notes and your full prompt text to your cmem.ai account. To keep everything local, skip the sign-in (CLAUDE_MEM_ONLINE_OPTIN=false). The observer is an extra AI working alongside every session: their hosted one during the trial, otherwise Claude on your plan by default, or Codex, Gemini, OpenRouter or any OpenAI-compatible endpoint. Notes are kept as written, so an outdated one sits beside the newer one; an opt-in setting added in v13.29.0 (CLAUDE_MEM_DEDUP_ENABLED, off by default) merges notes whose titles match. Anonymous usage stats are on by default; npx claude-mem telemetry disable turns them off. In January 2026 it wrote CLAUDE.md files into folders across people’s projects (#609, #641, closed as duplicates of #793, fixed in February); those files are now off unless you turn them on.
See what it stored: Its viewer, at /viewer on the address the worker prints when it starts.

claude-mem in the directory · Their site

Engram Alpha

Keeps a project’s reasoning: decisions and why they were made, problems and how they were solved, cautions, as typed notes joined by links like “replaces” and “because”. Tasks are kept beside the graph, not mixed into it. Your coding assistant writes the notes over MCP, and each session opens with a short brief. The memory is one file inside your repo, with no cloud and no keys. MIT. In Claude Code: /plugin marketplace add techtheist/engram.

What goes wrong: At 33 GitHub stars, it has few users so far. It saves only what the assistant decides to save. Duplicates merge when a note is saved; look-alikes and contradictions are only flagged by small local models, and you or the assistant decide. A replaced decision stays in history.
See what it stored: A graph you can read and edit, in your browser at 127.0.0.1:8787 or inside VS Code and JetBrains IDEs.

Engram Alpha in the directory · Their site

trusty-memory

A small local server that speaks MCP. The agent saves notes into a named “palace”, usually one per project, each embedded on your machine, beside a small graph of facts. trusty-memory setup adds a session-start hook and a per-prompt hook to Claude Code. When the machine is idle, a background pass merges near-copies and deletes old, unimportant notes. MIT.

What goes wrong: Its repo, which holds several tools, has 22 stars. Notes aren’t written for you: the agent or a hook has to save each one, and recalling a note doesn’t strengthen it. Its optional AI merge step needs an OpenRouter key or a local Ollama model.
See what it stored: Its MCP tools (memory_list, palace_info), or the dashboard in the separate trusty-console package.

trusty-memory in the directory · Their site

Mem0

A memory service, mainly for developers adding memory to their own apps, with a Claude Code plugin. Hooks capture your messages, Claude’s replies and the files that changed, and every five exchanges send them to Mem0, which pulls out short facts about the repo and about you. A new session recalls up to five. Mem0’s cloud merges duplicates and marks a contradicted fact as replaced; its Dream docs list this on every plan, while its pricing page lists Dream from Pro up.

What goes wrong: What it captures, your messages and Claude’s answers included, goes to Mem0’s hosted service; values that look like credentials are removed first, and whole files stay on your machine. It needs a Mem0 API key, and usage events are sent under your account email unless you set MEM0_TELEMETRY=false. The free tier covers 10,000 adds a month; paid plans start at $19.
See what it stored: /mem0:search, or ask Claude to search. /mem0:forget deletes your memories for the repo; shared project memory stays unless you add --include-project-memory.

Mem0 in the directory · Their site

Supermemory

A memory service that turns what you send it into short facts linked to what it already knows, so a new fact can replace or extend an old one. The plugin saves your conversations and important tool use, and before each turn Claude decides whether a memory search would help.

What goes wrong: By default it runs on Supermemory’s hosted API, with a key, where its own models do the extraction. To keep it on your machine, start npx supermemory local, point the plugin at it, and bring a model; the engine’s source isn’t public. In March 2026 the plugin was reported writing memories while search came back empty (#792, closed as fixed in July). The free plan includes the plugin with $5 of credits a month; Pro is $19.
See what it stored: /supermemory:session links to the current session’s document. Its docs don’t describe browsing everything stored from inside the plugin.

Supermemory in the directory · Their site

Cognee

Open-source memory that turns documents, code and chats into a knowledge graph. The plugin keeps your prompts, tool steps and answers during a session, adds relevant context to each prompt, and folds the session into the graph at the end. By default it starts a Cognee server on your machine (it needs Python 3.10 to 3.14).

What goes wrong: The local server still needs an AI to build the graph: your own model key, or the “Claude observer”, which runs those calls on your Claude subscription. When two facts disagree, an opt-in review in improve() marks each one current, superseded or conflicting, by the date you give each source; it’s off by default. Cognee Cloud’s first workspace is free with 10 million tokens, then $1 per million.
See what it stored: /cognee-memory:cognee-search, or ask a new session “What do you know from cognee?”

Cognee in the directory · Their site

Counterparts

Claude’s own memory of you and the relationship. At the end of each session Claude writes down what it learned, in its own words, and those memories fade unless they come up again. It keeps a journal of your time together and a self page: who it thinks it is, and who you are. It lives in ~/.counterparts/ as one SQLite database, plus a rebuildable search cache and daily snapshots, shared by every project on that machine; counterparts scope . --off keeps a directory out. No API keys and no account; search by meaning runs on a small model on your machine. Free, MIT. Install it from the Claude Code page.

What goes wrong: It uses your Claude plan too. Once a day (on by default for a new install; “no dreams” turns it off), a nightly run works in background claude -p calls: a dream, a reflection, and writing up any session that ended before it was written up. A typical night is about 250,000 tokens of new work; a night is capped at about 700,000, so a heavy day can take two nights. It isn’t a work memory, so keep your CLAUDE.md, which is also where a fact you must never lose belongs: fading is the design. Nothing syncs between machines. It’s early: its author’s daily memory on macOS since September 2026; Linux passes a smoke test but has no daily user yet; no Windows. The local search model misses a description that never names the thing.
See what it stored: counterparts dashboard (read-only, in your browser), counterparts self-page, or counterparts ask "…".

Counterparts in the directory · Their site

Three more directory systems reach Claude Code but aren’t built for this job. Zep’s Context MCP Server and plugin connect you to a memory an organisation keeps in Zep, with a Google Workspace or company login. Letta’s Claude Code plugin, Claude Subconscious, calls itself a demo “not intended to be used in production”; its README lists a Letta API key as required (LETTA_BASE_URL is offered for self-hosted servers). Hyperconsciousness, an encrypted store agents reach over MCP on expiring grants, calls itself a developer alpha.

Alternatives

claude-mem alternatives

Which one fits depends on why you’re looking.

The sign-in and the cloud. Signing in during install turns on sync, and your prompts go to cmem.ai. You can keep claude-mem and skip the sign-in, or move to Engram Alpha or trusty-memory, which have no account to sign in to.

The token cost of a second AI. Auto memory has no observer: Claude writes its notes inside the session you’re already in. Engram Alpha and trusty-memory work the same way, with your assistant saving notes through their tools.

Files appearing in your folders. claude-mem no longer writes folder CLAUDE.md files unless you turn them on. If you want nothing in the repo at all, auto memory and trusty-memory keep their files in your home folder; Engram Alpha keeps its one file inside the repo.

Wanting the memory in other tools too. Mem0 or Supermemory reach Cursor, Codex and your own apps, at the price of a cloud account (or, for Supermemory, a server you run).

Counterparts isn’t a claude-mem alternative: it keeps no notes on your code. For built-in memory and claude-mem side by side, see Claude Code vs claude-mem.

FAQ

Questions people ask

What’s the best memory for Claude Code?

For most coding work, the one it already has: CLAUDE.md plus auto memory, with claude-mem if you want every session written up for you. Which memory plugin should you use? above covers the rest.

Does Claude Code remember between sessions?

Partly: your CLAUDE.md files and auto memory notes carry over, the conversation itself doesn’t. Our guide to Claude Code’s built-in memory covers what loads and how to resume an old conversation.

Is there a local memory for Claude Code, with no cloud account?

Claude Code’s own memory, Engram Alpha, trusty-memory and Counterparts store everything on your machine with no account. claude-mem does too if you skip the sign-in, which otherwise turns on sync to cmem.ai. Supermemory can run as a local server; Mem0’s plugin uses Mem0’s cloud. claude-mem’s observer (Claude, by default), Cognee’s Claude observer and Counterparts’ nightly run call Claude on your plan, as your sessions do.

How do I see what my memory plugin stored?

Each card above has a “See what it stored” line. Built in, it’s /memory; every note is a markdown file.

Do I need a memory MCP server for Claude Code?

Usually not. If what you keep repeating is a rule, a line in CLAUDE.md fixes it; add a plugin or MCP server for something the built-ins don’t do.

Which Claude Code memory plugin scores best on benchmarks?

The tests vendors quote, LongMemEval and LoCoMo, ask questions about long chat histories rather than coding work, and vendors mostly run their own setups; Supermemory’s MemoryBench is an open harness for running one benchmark across several providers. Our benchmarks page shows how Counterparts does on those tests, with our setup named on every row and each vendor’s as it states it.

What about Honcho, Hindsight, MemPalace or basic-memory?

People compare those too. They aren’t in our directory yet, so we haven’t checked them the way we checked the systems here. Makers can add an entry by pull request.

References

Sources

Every maker’s docs and every issue read on 20 October 2026. These tools change every few weeks; if something here has gone stale, tell us.