Best memory tools for Claude Code: choose the missing job, not another memory layer

Claude Code already has memory. Before installing another tool, work out what it is failing to remember, who needs that information, and what should happen when it stops being true.

If your problem is repeating personal preferences, start with native auto memory. If you need stable project instructions, use a short CLAUDE.md or AGENTS.md file. If the gap is recalling past sessions, sharing knowledge across tools, or preserving an accepted product decision, an extra layer may help. Those are different jobs.

This comparison covers the native baseline, Supermemory, claude-mem, Hippo, Basic Memory, MemoryLake and Lockstep. It describes documented behavior, not a laboratory ranking. Lockstep is ours, and its section says where it fits and what remains unverified.

Start with the native baseline

Claude Code has two complementary mechanisms: instruction files you maintain, and auto memory Claude writes as it learns from corrections and preferences.

Project instructions can be shared through version control. Auto memory is machine-local, with one repository's worktrees sharing a memory directory. It does not automatically give a teammate on another machine the same knowledge.

That is a useful boundary, not a defect. Your preference for small commits need not become company policy. A product rule about who can export customer data should not live only in your personal notes.

Keep the baseline if it solves the problem. Another plugin adds installation, retrieval behavior, data retention and maintenance decisions. It should earn that complexity.

Supermemory: automatic recall with a shared storage layer

Supermemory's Claude Code plugin retrieves relevant memories before substantive prompts and captures new conversation content through hooks. Its integration supports a custom endpoint for self-hosted Supermemory as well as the hosted service.

It is worth evaluating when the missing job is recalling prior work across sessions and tools. Do not assume that a repository container means every teammate sees every captured item: check the organization, scope and access rules for your chosen setup.

The plugins are available on the free plan, but API usage consumes credits. At the 29 October check, Free includes $5 in monthly credits and one seat; Pro is $19 a month with $20 in credits and three seats. That is a usage model, not unlimited free memory.

Before using it on company work, inspect what the capture hooks send and who can retrieve it. A recalled conversation is useful evidence, but not proof that its tentative conclusion was accepted.

claude-mem: continuity from captured sessions

claude-mem captures tool-use observations, produces summaries and makes relevant context available to future sessions. Its current documentation also describes optional cloud sync and TeamBrain, so treating it as exclusively personal or exclusively local would be inaccurate.

It is a candidate when the gap is remembering work already performed: what was investigated, what failed and what the next session needs. Check its current installation instructions rather than assuming a global npm install configures the hooks; the README distinguishes the library from the working integration.

The limit to investigate is meaning. Can your setup distinguish an experiment from an accepted constraint? Can you find the source and retire a superseded conclusion? Session summaries alone do not answer those questions.

Hippo: a local-first store, with important network exceptions

Hippo stores memory in SQLite and writes readable Markdown mirrors. Its default BM25 search does not require an embedding model. Optional embeddings and other retrieval features change the setup.

Local-first does not mean every configuration keeps every operation offline. The README says its consolidation step sends memory text to Anthropic when an Anthropic API key is set. Optional API embedders and rerankers can also make network calls. Its setup can add hooks and a daily scheduled job, so review those changes rather than treating installation as passive storage.

It is a candidate when inspectable local storage matters and the team is comfortable owning the configuration. Test the intended configuration's network behavior before putting private project context into it.

Basic Memory is built around structured notes, links and a knowledge graph accessed through MCP. Its Claude Code integration includes a plugin with briefing and checkpoint behavior.

There is a meaningful distinction between local use and Teams. The local edition runs on your computer without team features. Teams adds a shared cloud knowledge base with per-seat billing, so one person's note can become available to another person and their agent.

This is worth evaluating when people should be able to read and maintain the same knowledge their agents retrieve. Shared notes still need ownership and review. A note recording a proposed plan should not silently become a rule simply because every agent can find it.

MemoryLake: evaluate shared project memory, not the marketing premise

MemoryLake describes project memory through MCP, shared access across teammates and role-based permissions. Its integration page describes free individual use and paid team collaboration.

Those are relevant capabilities to test. However, its claim that Claude Code is stateless and that memory must live outside the product should not guide the purchase: Claude Code's current official documentation describes native auto memory.

Evaluate the added value over that baseline. Test which events trigger capture, what loads in a new session, how permissions apply and what happens when a decision is replaced. This article has not independently tested those behaviors or verified a particular team's hosted setup.

Where Lockstep fits: accepted decisions across product and engineering

Consider an illustrative example. A product lead and engineering lead agree that exports must exclude a sensitive field. An engineer's next agent session sees an old prototype that included it. The missing fact is not merely a past conversation. It is which rule was accepted, why it exists, who owns it and whether it still applies.

Lockstep focuses on that decision record and the handoff to the people and agents doing the work. Its published CLI contains decision proposal, context and session-briefing operations. It also contains consent-aware advisory checks. A partial or unavailable check is not a clean result, and an advisory finding is not a guarantee that conflicting code cannot ship.

The individual onboarding path is currently a Claude Code pilot. The repository and website have some release-status differences, and this comparison has not verified a fresh authenticated hosted onboarding run. Treat the documented setup as something to validate on one repository before a wider rollout, not a promise about every integration.

Choose this category when the gap is accepted product and engineering choices surviving a session, a teammate or a tool. Do not add it simply to remember your preferred test command. Other products can store decisions too; compare the review, replacement and source-traceability workflow rather than claiming exclusive ownership of the word memory.

Test the missing job before adopting

Use a small test you can inspect, not a vendor's recall score:

Record an accepted decision, its reason and a rejected alternative. Close the session and ask a fresh one what applies.
Replace the decision. Check that the old version is not returned as current advice.
If sharing matters, repeat from a second account or tool with the intended permissions. Do not test using one shared admin identity.
Inspect captured content, deletion behavior and what leaves the machine. Local storage and external model processing are separate questions.
Inspect how much context is inserted and whether irrelevant history crowds out the task.

These are proposed evaluation steps, not tests completed for this article. The right result might be that your existing files and native memory already suffice.

A useful memory setup gives each kind of information a home. Keep instructions reviewable, personal learning personal, unfinished work temporary, and accepted team decisions traceable. Install another layer only for the gap that remains.

Sources checked 29 October 2026

For the wider context stack, see the tools for AI coding-agent context. For connections beyond memory, see MCP servers for engineering teams. For the decision-record workflow, see decision-log tools for product and engineering.