How do you give Claude Code persistent memory across sessions?

Every new Claude Code session starts amnesiac: yesterday's decisions, conventions, and dead ends are gone. Memory skills fix that by persisting context outside the conversation, and memory-merger leads the field at 12,568 installs (skills.sh registry).

Key takeaways

• memory-merger tops the memory skills at 12,568 installs and 36,327 GitHub stars (skills.sh registry, GitHub); the wider field is ranked on the best skills for AI agents pillar.
• LangChain's deep-agents-memory (11,965 installs) brings agent-framework memory patterns to skill form.
• remembering-conversations (8,882 installs) searches past session transcripts - episodic memory rather than curated notes.
• The agentmemory suite splits the job into verbs: remember (6,880 installs), recall (6,842), each a small composable skill.
• Built-in memory (CLAUDE.md and auto-memory) covers standing instructions; skills add structured, searchable recall on top - see Anthropic's Claude Code best practices.

Why does Claude Code forget everything between sessions?

Each session is a fresh context window; nothing carries over unless something writes it to disk. Anthropic's built-in answer is the CLAUDE.md file, which is read at session start and holds standing instructions. That covers rules, but not history: which refactor you abandoned last week, why the API uses seek pagination, what the failing test turned out to mean. Memory skills persist exactly that class of knowledge, which is why the category spans several distinct approaches across the AI agents category.

What is the best memory skill for Claude Code?

memory-merger from the github/awesome-copilot repo is the best pick at 12,568 installs and 36,327 GitHub stars (skills.sh registry, GitHub). Install it with:

npx skills add github/awesome-copilot

It consolidates accumulated memory files, merging duplicates and pruning stale entries so the persisted context stays small enough to load every session. Strong alternatives: deep-agents-memory (11,965 installs, skills.sh registry) if you want LangChain-style memory patterns, and para-memory-files (10,194 installs) for PARA-method file organization. The ranked field lives on the best skills for AI agents page.

How does episodic memory differ from curated memory?

Curated memory stores what you chose to write down; episodic memory stores what actually happened. remembering-conversations (8,882 installs, 433 stars, skills.sh registry, GitHub) takes the episodic route: it indexes past session transcripts so the agent can search "have we debugged this before?" and pull the earlier exchange. Curated approaches like memory-merger produce cleaner context but miss anything you forgot to record. Mature setups run both - curated notes for decisions, episodic search for everything else.

What about composable memory verbs?

The agentmemory suite by rohitg00 (24,810 GitHub stars on the repo) decomposes memory into single-purpose skills: remember writes a fact (6,880 installs), recall retrieves by query (6,842 installs), and session-history and forget round out the lifecycle at 6,840 and 6,820 installs respectively (skills.sh registry). The verb split matters for agent pipelines: a subagent can be granted recall without remember, keeping write access to long-term memory behind a gate.

How should memory skills work with CLAUDE.md?

Keep them layered, not competing. CLAUDE.md holds standing instructions that apply to every session - build commands, conventions, deploy policy. Memory skills hold accumulated knowledge that grows and needs pruning - decisions, gotchas, past investigations. The failure mode is dumping session history into CLAUDE.md until every session starts by reading a novel; memory-merger's consolidation pass (12,568 installs worth of users hitting the same problem, skills.sh registry) exists precisely because unpruned memory files bloat. Skillselion's catalog lists 60,702 skills (live catalog count); the memory shelf sits inside the AI agents category alongside the orchestration tools it composes with.

A complete starter stack

• memory-merger (12,568 installs) - consolidation so persisted memory stays loadable.
• remembering-conversations (8,882 installs) - episodic search over past sessions.
• remember + recall (6,880 and 6,842 installs) - composable write and read verbs for pipelines.
• deep-agents-memory (11,965 installs) - framework-grade patterns if you run LangChain-style agents.

Common pitfalls

• Writing everything to memory - unpruned memory files grow until loading them eats the context window the memory was meant to save.
• Storing secrets or tokens in memory files; they persist to disk and often to git.
• Running two curated-memory skills at once, which produces duplicate, drifting records of the same decisions.

Pick one curated skill, add episodic search, and prune monthly.

Common questions