remem: Local-first memory for Claude Code and OpenAI Codex

Stop re-explaining your project every new coding-agent session.

Language: English | 简体中文

remem automatically captures, distills, searches, and injects engineering memory across Claude Code and OpenAI Codex CLI sessions. Decisions, bug-fix rationale, project patterns, and preferences stay available through hooks, MCP, CLI, and a localhost REST API.

A new Claude Code session recalls the earlier root cause, commit, and open TODO with memory citations and no re-explaining.

What remem gives you

• Automatic session capture and background LLM distillation.
• Project-scoped recall across Claude Code and Codex using one local store.
• Searchable decisions, bug fixes, architecture notes, preferences, and raw session evidence.
• Source attribution, staleness labels, suppression, review queues, and injection audits.
• SQLite with SQLCipher encryption by default for fresh installs.
• MCP, CLI, and authenticated localhost REST access from one Rust runtime.

remem prioritizes memory quality. Automatic capture is the primary path; manual save_memory calls supplement it when a decision needs to be recorded immediately.

Install in five minutes

Homebrew

brew install majiayu000/tap/remem
"$(brew --prefix remem)/bin/remem" install --target codex

Use --target claude for Claude Code. --target all configures every known host, including Cursor where its v1 renderer is supported.

Standalone installer

curl -fsSL https://raw.githubusercontent.com/majiayu000/remem/main/install.sh | env REMEM_NO_CONFIG=1 sh
~/.local/bin/remem install --target codex

npm or Cargo

npm install -g @remem-ai/remem
# or
cargo install remem-ai --bin remem

remem install --target codex

Codex installation, detection, doctor, uninstall, transcript scanning and default native-memory import all honor CODEX_HOME; the default is ~/.codex. Set it to an absolute directory before running these commands to select another Codex profile. Empty or relative overrides fail visibly before installation writes. Unrelated profile files stay untouched.

GitHub Releases: prebuilt binaries for macOS and Linux on x64/arm64, with published checksums. Use one canonical remem executable on PATH; remem doctor warns when hooks and terminals resolve different copies.

For channel-specific upgrades, platform boundaries, PATH drift, and manual install notes, read the installation and upgrade guide. The broader documentation guide links plugin and operational material.

Verify the installation

Restart the selected coding agent, then run:

remem doctor
remem status
remem search "last decision"

A healthy Claude Code or Codex installation injects relevant project memory at SessionStart and queues durable session distillation at Stop. Codex also uses UserPromptSubmit to capture each prompt and surface compact optional memory candidates. remem doctor checks the schema, encryption key, database, hooks, MCP registration, worker, and common install-path drift. Its diagnostics do not create the data directory or write or rotate log files; findings and stderr warnings remain visible, and existing log health is inspected.

Repository contributors can verify duplicate SessionStart suppression with the isolated executable smoke fixture.

For a focused, read-only view of current-memory truth:

remem doctor truth --cwd .

Why wasn't something remembered?

Use one read-only report to follow capture, extraction, review, current validity, and the latest recorded context run:

remem doctor memory "original phrase" --cwd .
remem doctor memory --host codex-cli --source-root local \
  --project PROJECT_KEY --session-id SESSION_ID --json

Copy the exact selector from remem raw sessions --json. Use --injection-run-id RUN_ID to inspect a specific destination context emission. The source session and destination context run are distinct. Each evidence source is bounded at 20 rows; omitted evidence is marked. Terminal output shows retained source/status counts and three samples per source; --json includes all retained evidence. A missing match or per-item audit is unknown, not proof that capture failed or a guessed drop reason. Query discovery combines raw search with literal stored-text matches; try an original phrase or the exact session when a paraphrase finds no evidence. See the diagnosis contract.

Remem helps an Agent continue work; Refine helps people inspect and reuse knowledge; agent-sessions parses native source formats. Refine consumes Remem's existing session references, content fingerprints and commit/session links instead of collecting another transcript copy.

Host support

Capability	Claude Code	Codex CLI	Cursor v1
MCP memory tools	Yes	Yes	Yes on macOS/Linux
SessionStart injection	Yes	Yes	Not supported
Automatic session memory	Yes	Yes, Stop-based and low-noise	Not enabled by the v1 installer
Tool-event capture	Installed hooks	No high-frequency Bash hook by default	Runtime command exists; no installed hook
Compiled command-rule enforcement	Optional warn/block on Bash	Not supported	Not supported
Windows	Supported	Supported	Not supported

Cursor's v1 installer registers MCP only. The verified observe and summarize runtime commands exist, but remem install --target cursor does not install automatic capture hooks or SessionStart injection.

The repository also includes a Codex plugin wrapper. See plugins/remem/README.md for local plugin runtime and explicit hook activation instructions.

Why use remem alongside built-in memory

Built-in MEMORY.md, CLAUDE.md, and agent instruction files are ideal for a small set of stable facts that should always be visible. remem covers the engineering history that is too large, dynamic, or evidence-heavy to maintain by hand.

Need	Built-in files	remem
Stable project rules	Excellent	Supported
Evolving context	Handwritten instructions; Claude Code auto memory also writes notes	Hook-driven session capture and background distillation
Search older rationale	Limited by loaded text	Curated and raw search
Branch, time, and staleness handling	Manual	Built in
Provenance and injection audit	Git history	Database-backed audit
Review, suppression, and lifecycle governance	Manual edits	First-class commands

Claude Code auto memory already saves notes automatically. The distinction is the memory surface and its evidence, retrieval, and lifecycle controls, rather than automatic persistence alone.

Use both. Keep concise rules in native files and let remem retain the long tail of decisions, failures, evidence, and changing project state.

The broader ecosystem comparison lives in the dated memory-tool survey.

How it works

Claude Code / Codex hooks
          |
          v
append-only captured_events ledger
          |
          v
coalesced background extraction and session rollup
          |
          v
governed candidates -> curated memories + workstreams + raw archive
          |
          v
FTS, entity, temporal, vector, graph, and optional local rerank retrieval
          |
          v
budgeted, source-attributed SessionStart context

Hooks return quickly after durable capture or queueing. Background workers perform extraction, candidate governance, compression, retrieval enrichment, and lifecycle cleanup. MCP, CLI, REST, and SessionStart share the same local store and governance model, but apply surface-specific eligibility policies. Explicit search is an inspection and recovery surface, so it may return labeled legacy_unverified memories; default SessionStart and CurrentTruth exclude those rows and record the reason.

Malformed generated extraction responses retry with the existing bounded backoff while keeping the same source evidence. Invalid stored evidence still requires repair; exhausted model attempts preserve raw capture and replay ranges for inspection instead of accepting incomplete output.

Observation extraction and session rollup split large backlogs into bounded event chunks. Each model call budgets system instructions, prompt formatting, and evidence together; completion advances only through the processed chunk. Raw evidence remains available when per-event prompt content is clipped. Exact recovery retains the original range and resumes verified successful chunks under the explicitly selected profile and one overall timeout. Already-existing linked successors resume from their own original bounds and checkpoints; a completed primary task does not hide a failed successor. Archived exact recovery claims the existing family atomically, suppresses new follow-up work, and archives unfinished members again if the attempt fails.

Workers persist fair queue and stage/project claim order, so continuing capture in a busy project leaves room for downstream extraction and durable jobs across once-worker restarts. The existing once-worker admission budget still applies.

Generated memory is treated as untrusted until it passes source-support, secret, instruction-pattern, scope, and lifecycle checks. Unsafe content is dropped or routed to review with a diagnosable reason. Complete, user-supported security prohibitions, research activities and explicitly analytical quotations can enter human review in English or Chinese; they never auto-promote. Secret values, harmful instructions and unsupported external text remain excluded.

Automatically promoted summary decisions and discoveries can enter current context at their original summary confidence when captured local evidence, the current summary gate, and the exact activation receipt still validate. External content and missing or changed proof remain review-gated or excluded. New trusted evidence can reconsider identical content that is still pending review: remem preserves the original snapshot and links it to a fresh candidate. Repeated evidence, human review decisions, quarantine, and suppression are preserved. Expired operational state needs fresh evidence before renewal.

For module ownership and current data flow, read docs/ARCHITECTURE.md.

The experimental MCP context_bundle tool exposes the versioned, budgeted compiler to explicit callers. The experimental remem context-plan command prints a request-specific retrieval plan. These opt-in interfaces are tracked by the Context Bundle and retrieval-router contracts.

The default Codex integration stays low-noise: SessionStart provides stable context, UserPromptSubmit provides a compact candidate index, and Stop queues background summarization. Prompt candidates contain IDs, titles, state, retrieval reason, estimated read cost, and a detail lookup hint, but no memory bodies. They are optional leads that Codex may ignore, open, or search beyond. The first prompt may also receive up to two continuity anchors so a prompt such as continue does not depend on lexical overlap. Existing hybrid RRF ranks memory candidates; no final confidence threshold decides relevance for the model. Automatic extraction asks for short, specific candidate titles without a redundant project-name prefix; existing stored titles remain intact.

For your own Codex sidebar, an optional rename convention is MMDD｜TYPE｜Topic, such as 0903｜FIX｜FTS5 tokenizer mismatch. If you use a separate host-supported rename workflow, preview its proposed names before applying a batch. This is host-sidebar organization only: Remem does not rename host conversations or import those names as intent overrides. Remem's structured session/workstream fields remain authoritative, and a host title does not change memory types or workstream identity. See the session intent contract and implementation epic #1065.

Session/workstream label corrections use authenticated POST /api/v1/session-intent/preview, then explicit confirmation through POST /api/v1/session-intent/apply. Keep each sensitive option or Bearer prefix and its value on the same logical line in the proposed topic and reason. For example, curl --oauth2-bearer\nexample-value is rejected with HTTP 400 session_intent_cross_line_sensitive_argument; put the option and value on one line before retrying. Rejection creates no preview token or audit entry and changes no labels. Same-line sensitive values are redacted before preview.

Everyday workflows

Recall and inspect

remem search "database encryption"
remem search "deployment decision" --branch main --explain
remem show <memory-id>
remem why <memory-id>
remem current <state-key>

remem search keeps the terminal clean: per-query [INFO] [search-perf] diagnostics are written to the log file, not stderr, in normal use. Set REMEM_DEBUG=1 to mirror them to stderr while debugging.

Project aliases for Git worktrees

When a worktree has its own project key, preview an alias to the main checkout, then apply it after checking the paths and Git proof:

remem project alias add /path/to/worktree --canonical /path/to/main \
  --actor you --reason "same project"
remem project alias add /path/to/worktree --canonical /path/to/main \
  --actor you --reason "same project" --apply
remem project alias list
remem search "decision" --project /path/to/main
remem project alias revoke /path/to/worktree \
  --actor you --reason "worktree retired" --apply

The main checkout must already have a project row from normal remem capture. The alias keeps historical rows at their recorded paths while CLI and MCP project searches include the worktree's memories. add and revoke preview without database writes unless --apply is present. Revocation removes the worktree from the main project's search scope; list reports active aliases with the actor and reason from their latest activation.

Agents can use MCP search for compact results, then get_observations for selected details. Use raw recall only when curated memory misses exact transcript evidence:

remem raw search "exact phrase" --since 2026-06-01 --json

List complete host-bound sessions before reading an exact transcript:

remem raw sessions --latest 20 --json
remem raw messages --host codex-cli --source-root local \
  --project "/path/to/project" --session-id SESSION_ID --json
remem ingest-sessions --root codex-cli:archive=/path/to/sessions --json

Default transcript scans use $CLAUDE_CONFIG_DIR/projects and $CODEX_HOME/sessions, falling back to ~/.claude and ~/.codex. An explicitly empty override is an error. Explicit override directories must exist and be readable; only absent inferred defaults are skipped. Explicit --root directories remain supported; descendant subagents directories are excluded from batch scans. Codex session modes use native source evidence: IDE is interactive, explicit exec is unattended even with a Desktop originator, and subagent evidence takes precedence. These are source labels, not proof of who initiated a run. Existing saved modes upgrade once after checking the original classification evidence; genuine provenance conflicts still fail the batch. This adds schema v093: retain a database backup before upgrading; older binaries cannot open the upgraded schema. Raw message identities and cursors are preserved.

Copy host, source_root, project, and session_id unchanged from one raw sessions summary into raw messages. Existing scripts must add the required --host selector and replace --root LABEL=PATH with --root HOST:LABEL=PATH; the same root format applies to raw reconcile. The JSON envelope reports excluded_legacy_rows, excluded_legacy_sessions, and excluded_legacy_identities when some archive rows cannot enter the host-bound session contract. Each listed session also carries additive nullable mmdd, session_intent, session_topic, display_label, and session_intent_source fields. mmdd is derived from the session created epoch in Asia/Shanghai; the full {MMDD}｜{INTENT}｜{topic} label is present only when both intent and topic are known. Stop summaries may fill those fields with source summary; unknown, empty, too-long, or redacted values abstain instead of blocking the rest of the summary. Automatic Stop rollups preserve manual label overrides, including explicit clears, across later turns. Listing still returns healthy sessions; --latest N fills that bound from healthy sessions only and does not let unresolved rows occupy those slots. Use the skipped identities (source_root, project, session_id, and host when known) to inspect or repair those rows. Exact raw messages for a skipped selector stays fail-closed. Do not re-ingest a skipped row unless ingest can actually claim it — many legacy rows have no trusted host provenance. HOST is claude-code or codex-cli, and LABEL becomes the persisted source_root. Cursor snapshot evidence requires a manually configured and verified remem summarize --host cursor Stop integration; filesystem --root ingestion and reconciliation reject cursor explicitly.

Review and govern

remem review list
remem review approve <candidate-id>
remem memory suppress memory:<id> --reason "no longer relevant"
remem govern --action stale --dry-run --json <id>

Mutating governance commands expose previews, explicit confirmations, or review boundaries according to their risk. Run remem <command> --help for the current contract instead of relying on a copied command inventory.

MCP tools share that store with stricter wire contracts. The canonical MCP contract is GH981, including the #1061 mutation and scope boundary:

• save_memory: pass host when the calling host is known. An omitted host is recorded as unknown, never inferred as codex-cli.
• govern_memory: dry-run first to preview IDs and current versions from that governance transaction. Non-dry-run mutations require expected_versions for every ID, confirm_destructive=true, and an explicit reason.
• recall_user_context: supply project or cwd. The server does not infer this scope from its own process working directory.

The local Remem app shows session/workstream labels or explicit abstention. Sessions can be filtered by intent and created date in Asia/Shanghai. Operators can correct labels in batches with a Before/After preview, explicit confirmation, and an audit; workstream identity and host conversation titles stay stable.

Configure memory AI and retrieval

remem config show
remem model current
remem model use balanced --dry-run
remem embedding status
remem embedding download --model multilingual-e5-small
remem embedding backfill --limit 1000

auto embedding mode stays local unless a remem-specific API key is selected. The verified local model is optional; the labeled feature-hash fallback remains available. The second-stage local reranker is also optional and disabled until configured.

Vector indexes keep each model and artifact revision in its own vector space. Project, branch, type, and lifecycle filters apply before nearest-neighbor selection. Automatic context uses the same vector executor as search, so old memories remain eligible by similarity as a project grows. While a derived index is rebuilding, an exact local scan preserves recall; large stores may take longer until that profile's index is ready.

Use the current configuration routes, the local embedding contract, and remem config, remem embedding, or remem reranker help for details.

remem usage reports every dispatched memory-AI attempt, including failures. Observed zero, partial, missing, invalid, and text-estimated token counts remain distinct. Reported USD is the known priced portion; coverage counts show when unknown rates or missing counter details make it incomplete. Failed attempts without provider counters do not receive guessed token estimates. Historical usage remains labeled as unverified rather than being silently reclassified.

Share or edit memory outside the database

remem sync-memory --cwd .
remem export --markdown --output ./remem-memory
remem export --pack .remem-pack

Markdown mirrors are human-editable. Project memory packs are deterministic, git-committable exports with provenance-aware import and quarantine behavior. See the memory usage guide and project memory pack contract.

Evidence and benchmarks

The checked-in public suite separates memory-system capability evidence from coding-agent outcome evidence. Verify it locally with:

cargo run -- bench verify --root eval/public --json-out /tmp/remem-bench-verify.json

Verification resolves claims/registry.json beside the parent of --root, so an external bundle keeps public/ and claims/ as siblings and is independent of the caller's working directory.

Public adversarial SQLite snapshots are capped at 64 MiB and must be canonical VACUUM images; verifier-consumed artifact targets must also resolve inside the declared public root.

The current public report does not support public benchmark claims and is deliberately labeled directional_only_no_public_claim. The historical isolated coding baseline is useful engineering evidence, but its preloaded-memory condition is not comparable with the current SessionStart retrieval path.

Reproduction commands, artifact schemas, claim boundaries, and current gates live in:

• eval/README.md
• eval/public/README.md
• eval/coding-bench/README.md

README claims intentionally exclude unsealed local metrics that have no checked-in report.

Security and privacy