they gave you a sticky noteand called it memory.
CLAUDE.md is a text file your AI re-reads top to bottom, every turn. ImmorTerm Memory digests every session as it happens and hands your AI the right memory in ~8ms — before you ask.
one binary · no api key · nothing leaves the machine
Rotate refresh tokens server-side; access tokens stay 15 min. Replaced the sliding-window plan from feature/auth-spike.
linked: auth/refresh.ts · commit 4f2a91c · session 2026-06-30
a filing job, wearing infrastructure's clothes.
You install the plugin. You design the folder structure. You hand-write the naming conventions. You tell the AI to summarize at the end of every session and read the logs at the start. You review its backlinks one by one. And every week you prune, because past ~500 notes retrieval falls apart and the AI starts confidently quoting decisions you reversed a month ago.
It's a filing job you inherited from your own tools. ImmorTerm Memory does the filing — ambiently, forever.
session → digest → recall. hands-free.
no notes written. no folders organized. no pruning. zero maintenance.
it ages with your repo. not against it.
ambient digest. you write nothing.
The filter (LangGraph pipeline + Claude extraction) reads every session as it happens and files the decisions, root causes, and changes. No end-of-session ritual, no summaries to author, zero integration code. Mem0, Zep, and agentmemory are libraries your agent has to remember to call — this captures whether anyone asks or not.
it supersedes itself. no scars.
Your repo changes its mind; your memory should too. When a new decision contradicts an old one (cosine similarity + an LLM verdict), the old memory is superseded — chained for forensics, no longer surfaced as current truth. The record matures with the project instead of rotting against it.
branch-aware recall
Memory is scored against your current git branch: off-branch memories take a 0.85× penalty, and merged-scope memories drop out. Move from feature/x to main and the ranking re-sorts itself. Your AI stops suggesting patterns from an archived branch.
code-bound memories
Every fact is linked to the file, diff, commit, session, and checkpoint it came from. Reconstruct a file at any past version, walk a supersession chain, or explain why a change happened — a full forensic trail, not plain text. Writes go to the database only; your repo files are never touched.
recall injected before you ask
Relevant context surfaces into the next session automatically — you start work already caught up, instead of re-explaining last week to your AI. When you want to dig, there are 41 MCP tools: filter by date, category, session, branch, file, or entity; inspect, edit, or delete anything it remembers.
three commands. then the filing starts.
The installer drops a single native binary and wires the MCP config for Claude Code. No Docker. No API key.
macOS (Apple Silicon) and Linux (glibc 2.38+ — Ubuntu 24.04 / Debian 13 or newer).
Recall is automatic. The tools are for the days you want the forensic trail: filter by date, branch, file, or entity; walk a supersession chain; rebuild a file at any past version.
Five ONNX models in one Rust binary — embedder, NER, cross-encoder reranker (ms-marco-TinyBERT), intent classifier, and injection gate. Hybrid dense + BM25 retrieval, WAL-durable SQLite, ~8ms per query. No cloud, no keys.
questions, answered.
How is this different from Mem0 or Zep?+
Those are memory libraries your agent has to call, usually cloud-hosted with an API key and 100–500ms round trips. ImmorTerm Memory is local infrastructure that captures automatically and recalls in ~8ms — plus supersession, branch-awareness, and code-binding that library APIs don't ship.
Isn't CLAUDE.md already memory?+
CLAUDE.md is a static file re-read in full every turn — no ranking, no forgetting, no supersession. It can't answer "what's relevant now." This can.
Will it overwrite my notes or repo files?+
No. All writes go to a local database. Your files and markdown are never touched.
Does anything leave my machine?+
No. Search, embeddings, reranking, and classification all run on-device. No API key, no external calls for recall. Telemetry is opt-in and never includes your code, prompts, or memory contents.
What happens to old memories on Free?+
They're stored forever — just held past the 72-hour window until you upgrade. Nothing is purged. The gate is a window, never a purge.
Does it work outside Claude Code?+
Claude Code today, with the deepest integration. Cursor, Codex, Windsurf, Cline, and Aider are coming soon.
free keeps every memory forever. the gate is a window.
Local, yours, forever.
- Ambient digest, supersession, branch-aware scoring
- Code-bound memories
- All 41 MCP tools, all hooks
- 5 memory search results per query
- 72h memory retention
- Every memory stored forever on-device — nothing is deleted
For projects that outlive a week.
- Unlimited search results
- Unlimited retention — your entire history stays searchable
- Everything in Free
The free tier is gated, not crippled: every memory is preserved locally. Past the window they stop surfacing until you upgrade — and gated searches tell you exactly how many more matches are waiting. Or take everything — terminal, memory, and all 23 themes — with ImmorTerm Pro, $29/mo.
the filing job ends here.
the record doesn't.
local-first · ~8ms recall · claude code native · free is real