에이전트 강화
게시자가 제공하는 설치 방식입니다. PlugMCP 엔드포인트가 아닙니다.
원격 URL인가요, 내 컴퓨터에서 실행하는 명령인가요? 두 설치 방식의 차이
PlugMCP에서 실행
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
실행게시자 제공
Claude Code's memory is scoped per project directory. Your history isn't.
Every session you've ever run is already on disk in — every decision, every command that worked, every credential you pasted. But open a session in a different directory and none of it is reachable. You re-ask questions you already answered, and re-implement integrations you already wrote.
Turn your Claude Code session history into a searchable local knowledge base — 13 MCP tools, pure stdlib, nothing leaves your machine
~/.claude/projects/**/*.jsonlThis indexes all of it and hands Claude 13 MCP tools to search it. You just ask normally.
What was the retry logic in that deploy script? Which DATABASE_URL did I actually use for staging? Before I write the Stripe integration — has anyone here already done payment auth?
Pure Python standard library. No third-party dependencies, no embeddings, no API calls, nothing leaves your machine.
git clone https://github.com/nameforjt-afk/session-knowledge.git
cd session-knowledge
bash install.sh
Restart Claude Code. Done.
The index lives in ~/.claude/session-index/ and it contains plaintext credentials
scraped from your sessions. That is the precondition for "which key did I use here"
to work at all.
Never commit that directory, never sync it to cloud storage, never send it to anyone.
This repo's .gitignore covers the default location, but if you relocate it that's on you.
The full-text index itself is redacted: common credential shapes (API keys, Bearer
tokens, JWTs, GitHub PATs, passwords inside connection strings) are replaced with
⟦SECRET:fingerprint⟧ before they are written. Plaintext lives only in vault.db
at mode 0600, and only an explicit creds get returns it.
The transcript and code indexes still contain sensitive local context even after redaction. The index directory is restricted to mode 0700; every generated SQLite database, WAL/SHM sidecar, and refresh log is restricted to mode 0600.
Redaction is pattern matching, not magic. Custom-format secrets can slip through.
verify-redaction spot-checks for leaks.
Four things, in rough order of how often they save you:
1. "Have I already built this?" — Search by external service, not by function name.
grep only works if you already know the string to search for (you need to know it's
tenant_access_token before you can grep for Feishu auth). Capability tags don't
require that.
2. "How did we decide this?" — The rationale is buried in thousands of past instructions. Full-text search over them, filterable to just your own instructions.
3. "How did that command go again?" — Every Bash and MCP call, with arguments and results. Filter to only the ones that failed, to recall the traps.
4. "Which value does this key take?" — Same variable name routinely has several real values (multiple apps, dev vs prod, two tables). It returns all candidates ranked, with the call site each came from, rather than silently picking one.
A hook refreshes the index incrementally on each Claude Code start (async, non-blocking, usually 1–2s). Ask questions normally; Claude reaches for the tools on its own. When an upgrade changes parsing, redaction, or tokenization, the next refresh performs one automatic full rebuild and then returns to incremental updates. Code parser and capability-rule upgrades likewise trigger one automatic full code-index rebuild, preventing unchanged source files from retaining stale symbols or tags.
Deleting a transcript removes its searchable text, tool calls, metadata, and session-only credential observations on the next refresh. Credentials that still occur in another live transcript remain available.
python3 -m sessionmcp.cli index # incremental refresh
python3 -m sessionmcp.cli stats # index health check
python3 -m sessionmcp.cli search "deploy timeout" # full text (multi-word = AND)
python3 -m sessionmcp.cli search "pricing" --kind user_instruction
python3 -m sessionmcp.cli tool "docker build" # past commands and API calls
python3 -m sessionmcp.cli tool "stripe" --errors # only the ones that failed
python3 -m sessionmcp.cli timeline "that migration" # reconstruct how it unfolded
python3 -m sessionmcp.cli synth "rate limiting" # cross-session material pack
python3 -m sessionmcp.cli evolution "auth design" # how a decision changed over time
python3 -m sessionmcp.cli code find --capability stripe # who implemented payment auth
python3 -m sessionmcp.cli code find send_message # by symbol name
python3 -m sessionmcp.cli code dup # duplicate-implementation report
python3 -m sessionmcp.cli creds list # variable names only, never values
python3 -m sessionmcp.cli creds get DATABASE_URL # all candidates, ranked
python3 -m sessionmcp.cli creds get X --masked # masked form only
python3 -m sessionmcp.cli verify-redaction # confirm no plaintext in the index
Search is AND. More words means fewer hits. Use the 2–3 words you would actually have typed back then, not a sentence.
Not every "credential" in your history is real. Doc examples, error snippets, values someone made up mid-discussion — all get extracted, and they look exactly like the real thing. A fake in your candidate list is worse than no record at all: it looks perfectly plausible, so you use it.
Blacklist it permanently:
python3 -m sessionmcp.cli creds forget --fingerprint <fp> --reason "doc example"
Deleting alone does nothing — the index re-scans daily and the value comes right back.
forget is delete plus blacklist.
~/.claude/projects/**/*.jsonl transcripts Claude Code already writes
↓ single pass
┌────────┴────────┐
index side credential side
redacted → FTS5 KEY=VALUE → vault.db (0600)
↓ ↓
index.db vault.db code.db
full text credential symbols +
registry capability tags
CJK search uses bigram pre-tokenization. FTS5's built-in trigram cannot match two-character
Chinese words (measured: common two-char terms returned nothing), and unicode61 doesn't
segment CJK at all. So CJK text is expanded into adjacent character pairs at write time and
the query is expanded the same way — exact matches work, still on the FTS5 index, BM25
ranking preserved.
Most of it needs none. To tune, edit sessionmcp/config.py:
Auto-derived: working directories with ≥2 sessions in the last 45 days, top 30 by session count. Not hard-coding a list is deliberate — a hard-coded list goes stale silently as projects come and go, and is entirely wrong on a new machine.
If you're new to Claude Code and every project has exactly one session, this derives an empty list and the code index scans nothing. Override:
export SESSION_KNOWLEDGE_PROJECT_DIRS="~/proj-a:~/proj-b"
~/.claude/knowledge/canonical.json lets you declare which file is the canonical
implementation of a capability. The generator reads it and never overwrites it, so daily
refreshes won't clobber your call. When the same logic shows up a third time, claiming
one beats copying it a fourth.
bash uninstall.sh # remove MCP + hook, keep the index
bash uninstall.sh --purge # remove the index too
~/.claude/projects/**/*.jsonl). Not Cursor, not Copilot.verify-redaction periodically..py .js .ts .tsx .mjs .jsx).brew install python.See CONTRIBUTING.md before opening an issue or pull request. Never include real transcripts or credentials in public reports. Potential vulnerabilities should be reported privately according to SECURITY.md.
Release history is recorded in CHANGELOG.md.
MIT — see LICENSE.
Turn your Claude Code session history into a searchable local knowledge base — 13 MCP tools, pure stdlib, nothing leaves your machine
도구
아직 도구 목록이 없습니다. 원격 서버는 첫 상태 확인 후에 도구 목록이 생깁니다.
FAQ
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Session Knowledge은(는) stdio를 통해 로컬에서 실행됩니다.
마지막 확인: 2026-10-26. Turn your Claude Code session history into a searchable local knowledge base — 13 MCP tools, pure stdlib, nothing leaves your machine
