Session amnesia
Each new Claude Code or Codex session starts without the context you already earned.
Local-first memory for Claude Code and OpenAI Codex
Stop re-explaining your project every new session.
An open-source Rust CLI and MCP memory server that automatically captures, searches, audits, injects, and verifies local project memory across Claude Code, OpenAI Codex, and Codex CLI sessions.
- Single Rust binary
- Localhost only
- SQLCipher at rest
- MIT licensed
The problem
Every new session starts from zero.
Context you spent hours building disappears the moment the thread ends.
Lost rationale
Design decisions and bug-fix root causes vanish when the session ends.
Preference fatigue
The same project preferences get repeated instead of carried forward.
No continuity
Long-running work is hard to resume with confidence after context compaction or a fresh session.
How it works
Hooks into your agents. Nothing to capture by hand.
remem records raw session evidence, distills it in the background, then serves compact memory back through hooks and MCP.
Claude Code
SessionStart
Inject relevant project memory and preferences.
Claude Code
UserPromptSubmit
Register sessions and flush stale queues.
Claude Code
PostToolUse
Queue tool observations without blocking the workflow.
Claude Code / OpenAI Codex
Stop
Summarize and promote durable memories in the background.
Search architecture
Four retrieval channels, fused.
A query is routed through BM25, entities, temporal parsing, and LIKE fallback, then merged with Reciprocal Rank Fusion.
- Project-scoped search prevents cross-project leakage.
- CJK segmentation and Chinese-English synonym expansion improve recall.
- Title-weighted BM25 and topic-key dedup keep results compact.
"database encryption"
FTS5 BM25
Entity index
Temporal
LIKE fallback
RRF = sum 1 / (60 + rank)
Top-K merged memories
Product proof
The memory arrives before the work begins.
Current releases ship through GitHub, crates.io, npm, and Homebrew, with reproducible artifact checks guarding public benchmark and claim wording.
Benchmarks
Measured, not hand-waved.
Public artifacts now separate memory-system capability evidence from coding-agent outcome claims.
4 manifests, 4 reports, and 25 run artifacts verified.
Temporal answers, stale decisions, conflicts, file anchors, and user-context relevance.
Secrets, credentials, unsafe claims, stale anchors, and unresolved conflicts.
No SOTA or coding-task superiority claim until the public gate passes.
Usage & cost
Measured locally, reported as estimates.
remem records AI usage for its own background extraction, summaries, compression, and promotion calls. The CLI reports token usage and estimated USD cost by day, week, source, project, and executor.
- Provider/log usage is separated from text-estimated fallback rows.
- Cost is an estimate, not a provider invoice or subscription bill.
- Executor and model selection live in ~/.remem/config.toml.
- The local audit below is a snapshot from 2026-06-06; usage varies by workload.
remem usage --days 14 --weeks 8
Daily and weekly totals
Project-scoped reports
Usage source precision
Estimated USD cost
provider/log rows stay separate from estimates
cost estimate, not invoice
This repository's remem ledger was $6.18 against $17,595.10 of total agent spend.
All remem background memory calls across projects were $437.15 in the same window.
Estimated background memory cost for this repository over 8 weeks.
ccstats all-source spend over the same 2026-04-12 to 2026-06-06 window.
Install
Use the same commands as the README.
Start with Homebrew, npm, the prebuilt GitHub Release installer, or Cargo. Then restart your coding agent.
curl -fsSL https://raw.githubusercontent.com/majiayu000/remem/main/install.sh | shGive Claude Code and OpenAI Codex a local memory layer.
Single binary. Local store. Searchable, auditable context for long-running engineering work.