Memory Layer MCP

Most memory tools remember conversations. Most code-search tools index code. Memory Layer MCP does both — plus commits and docs — so Claude Code and Codex can answer not just "where is this code?" but "why was it built this way?"

One memory. Claude Code or Codex — doesn't matter. Questions you ask in one agent become context the other agent sees tomorrow. Auto-save, auto-recall. Cross-language code search (tested heavily on Korean ↔ English codebases).

Local by default:

• Q&A, memory, and indexes stay on your machine (.hybrid-search/, auto-gitignored)
• Secret-shaped prompts are dropped; credential-shaped files never indexed
• Only chunk text sent for embedding (OpenAI API) — no telemetry, nothing else
• Kill switch: HYBRID_SEARCH_QA_LOG=0

(Repo is named hybrid-search-mcp for historical reasons; the product and PyPI package are Memory Layer MCP / memory-layer-mcp. Both work as the CLI command.)

Real output, June feature on a 2,100-file production app: the conversations where it was decided (1–3), the working Q&A from build day (4–7), the merge commit that shipped it (8), and the March design doc (9) — one query. Asked in Korean, found across an English codebase.

Day 1 (Claude Code): "portal-v3 인증 흐름이 어떻게 되지?"
                     → answers from code, saves Q&A to .hybrid-search/qa/

Day 2 (Codex):       "portal 인증 어디서 처리해?"
                     → pre-fetch surfaces Day 1's exchange before Codex even searches

The Day-2 turn didn't re-search the code fresh — yesterday's Claude Code answer surfaced as context inside Codex. Every answered query becomes a first-class search result for every future query, in either agent.

Trade-offs you should know up-front (we've measured them):

• First-query latency adds ~400 ms of pre-fetch overhead (vs ~50 ms grep). Worth it for exploratory questions; not for grep-shaped lookups (and the router knows the difference).
• Embedder = OpenAI text-embedding-3-small (API key required). No local embedding backend — by choice, not neglect. The first version ran local models; bulk-embedding tens of thousands of chunks pinned an M3 MacBook's fans for the entire run and made the machine unusable (CPU path was no better). A full reindex of a 2,000-file project via the API costs cents. If zero-API-key is a hard constraint, this tool isn't for you today; a backend config field is reserved and a local ONNX contribution is welcome.
• This is not "0-config", and we don't claim it. Explicit prerequisites: macOS/Linux with Python 3.11+ and pipx or uv already installed, plus an OPENAI_API_KEY. From there it's one install + setup (Claude) + setup --codex (Codex). A stock Mac without those prerequisites will fail at install — see Requirements above for the misleading pip error this produces.

Who this is for: 1인 개발자가 Claude Code를 주력으로 + 가끔 Codex도 쓰면서, 같은 코드베이스에서 반복 질문을 줄이고 싶은 사람. Korean + English 코드베이스에서 검증됨 (valuein_homepage 708-commit, 1,307 files).

v0.3.0 (2026-04-23): deterministic guarantees. Four Claude Code hooks (PreToolUse, SessionStart, UserPromptSubmit, Stop) wire the memory layer into every turn — save is independent of Claude's tool choice, retrieval fires before Claude sees exploratory prompts. No stochastic "sometimes works" behaviour; every turn persists, every exploratory prompt gets pre-enriched.

On a 20-query benchmark against valuein_homepage, memory surfaces past Q&A in 80% of repeated queries and 50% of reworded follow-ups, lifting end-user "answer found in top-10" from 75% → 90% (paraphrase) / 80% → 90% (identity). Guaranteed-save is unit-test verified (TestStopHook suite) since the bench drives the orchestrator directly without hooks in the loop. See Compounding benchmark.

---

Why this is different

Most code-search and memory tools either (a) index source once and forget your conversations, or (b) live inside one agent's UI:

• Sourcegraph / Cody: static embeddings of source files.
• Cursor / Aider: ephemeral context, tied to one tool, forgotten next session.
• Graphify: knowledge graph, rebuilt on commit.
• ChatGPT Memory: personal preferences, no code context.
• Mem0 / Letta: agent memory, but you wire it in and manage facts manually.

Two things make this stack different:

1. Closed loop, automatic. The answer to your last question becomes indexed context for the next question. No "save this" command, no memory dashboard — every turn persists via the Stop hook, every prompt gets enriched via the UserPromptSubmit hook.
2. One memory, two agents. Claude Code and Codex hooks both read and write the same .hybrid-search/qa/ directory. Yesterday's Claude session informs today's Codex session, and vice versa.

Your .hybrid-search/qa/ directory is the log of every exchange, in plain markdown, grep-able and git-able.

The Memory Layer at a glance

When you…	What happens
Run hybrid_search	Query + top-10 results saved to .hybrid-search/qa/YYYY/MM/*.md
Run git commit	post-commit hook reindexes changed files and the commit message itself (with changed-file anchors)
Ask "how was X built / why did X change"	History-intent detection pulls past conversations, plan docs, and the commits that changed it into one answer
Ask "who calls X / where is X used"	A call-graph card (callers + callees with file:line) rides along with the top code hit
Want to see what's been captured	hybrid-search-mcp viewer --open — local single-file dashboard, nothing leaves the machine
Ask a related question later	Past qa logs compete for top-10 like any chunk
Say "지난번에…" or "previously…"	Memory-intent detection → 2× boost on qa logs
Let time pass	30-day half-life decay — stale answers quietly fade
Ask again after the fact changed	Same-topic supersession: the newest answer takes the earlier slot — 6/6 on the Korean dev set AND 6/6 synthetic (incl. Korean→English probes) on an untouched English OSS holdout (memory bench v2)
Ask about something the project never had	Confidence contract refuses instead of bluffing — 0 strong on nine verified-absent probes (8 weak, 1 mixed)
Uninstall	pip installs: hybrid-search-mcp teardown. Plugin installs: /memory-layer:teardown (the CLI lives in the plugin's own venv, not on PATH), then /plugin uninstall memory-layer@curiohunter
Paste a secret by accident	Regex filter drops sensitive queries before they touch disk

Turn it off anytime: export HYBRID_SEARCH_QA_LOG=0.

How it compares (honestly)

Star counts verified via GitHub API on 2026-07-10. ⚠️ means "claim is partly true." We won't pretend everything is ✅ — several of these projects are excellent and far more popular than this one.

Project (⭐)	Conversations auto-captured	Code index	Git commits indexed	Codex + Claude Code	한↔영 code search	No vector-DB signup
claude-mem (86.7k)	✅ hooks	❌	❌	✅ many	❌	✅
agentmemory (24.9k)	✅ hooks	❌	❌	✅ many	❌	✅
mem0 / OpenMemory (60k)	⚠️ agent-decides	❌	❌	⚠️	❌	⚠️ cloud-leaning
claude-context (12.1k)	❌	✅ hybrid	❌	✅	❌	❌ Zilliz/Milvus required
Serena (26.3k)	❌	⚠️ LSP symbols	❌	✅	❌	✅
memory-layer-mcp	✅ hooks + per-turn index	✅ hybrid BM25+vector	✅	✅ transcript-level	✅	✅ (embedding API key only)

What this table says, plainly:

• The one thing nobody else does: one index over your code + conversations + commit messages + docs, so "how was this feature built?" returns the discussion, the plan, the commit, and the code in a single query. Conversation-memory tools stop at conversations; code-index tools stop at code; git-search tools stop at git.
• Where others are ahead: claude-mem/agentmemory have far more users and integrations; Serena adds symbol-level editing we don't do — install both; they compose well (Serena edits symbols, this remembers why); claude-context scales to multi-million-LOC monorepos.
• Honest concessions: OpenAI embedding key required (no local backend — tried, laptop couldn't take the bulk load); pre-fetch adds ~400 ms per prompt; and for exact-symbol lookups plain grep is still faster — our router sends those to grep on purpose.
• Numbers you can re-run, not vendor slides: stale-fact supersession 6/6 on the Korean dev set and 6/6 synthetic on an untouched English OSS holdout (ripgrep, single run, published as-is — CHANGELOG-derived planted cases 3/5 with the misses diagnosed), zero false-strong across 27 verified-absent probes on three codebases, ≈ 3 k tokens per answer — scripts in benchmarks/, failures published alongside the wins (see Memory bench v2).

Built-in guardrails (the parts that make automatic capture safe): secret-shaped queries never touch disk; files that look like credentials are never indexed; junk turns are gated and re-asked questions aren't stored twice; memories fade on a 30-day half-life instead of going stale silently; every result carries provenance (trust_meta) and each response reports how much of it is self-generated content (generated_ratio) — our retrieval-collapse early warning. Inspect everything captured at any time: hybrid-search-mcp viewer --open.

---

Privacy & Data

Read this before installing — the Memory Layer stores conversation-derived content on disk by default, because that is what makes the compounding loop work. Everything below is local, plain-text, and opt-out.

What is stored, where:

Data	Location	Default
Q&A logs (your query + answer excerpt + top results)	<project>/.hybrid-search/qa/	on
Memory cards / facts	<project>/.hybrid-search/memory/	on
Generated wiki pages	<project>/.hybrid-search/wiki/	on
Indexed agent transcripts (Claude Code + Codex turns)	global index, from index-conversations / per-turn hooks	on when hooks installed

What leaves your machine: chunk text (code, docs, qa logs, conversation turns) is sent to the OpenAI embeddings API for embedding — nothing else. No telemetry, no other network calls. There is currently no local embedding backend — the first version ran locally, and bulk-embedding a real codebase on laptop hardware proved unusable (sustained fan-pinning load on an M3 MacBook; CPU path no better), so the API is a deliberate trade-off. If sending chunk text to OpenAI is unacceptable, do not install this tool yet; a local ONNX contribution is welcome ([embedding] backend field is reserved for it).

Safety rails built in:

• A sensitive-query regex drops password/token/secret-shaped queries before they ever touch disk; files that look like secrets are never indexed.
• A quality gate drops junk turns; a re-asked question within 7 days is not stored twice.
• Indexed conversation turns containing prompt-injection-shaped text are tagged with an [untrusted content …] banner at display time.
• setup writes .gitignore entries so .hybrid-search/qa/ and memory never end up in your repo — teammates don't see your conversation log unless you deliberately commit it.
• Sensitive subfolders (patient records, contracts, HR docs, …) can be kept out of indexing entirely: list them in a .hybrid-search-ignore file at the project root (gitignore syntax) and re-run hybrid-search-mcp index . --force. Excluded files are never chunked, never embedded, never sent anywhere.
• Retention: 90-day / 2,000-file auto-prune with a dry-run first pass; archived entries live 30 more days and are restorable (qa-restore).

Full opt-out:

export HYBRID_SEARCH_QA_LOG=0     # stop persisting Q&A turns
export HYBRID_SEARCH_INDEX_QA=0   # stop indexing existing qa logs
export HYBRID_SEARCH_ROUTER=0     # stop per-prompt pre-fetch injection

---

Quick Start

Requirements

• Python 3.11+ — heads-up: on a stock Mac the system pip3 runs Python 3.9 and fails with a misleading resolver error (e.g. Could not find a version that satisfies tree-sitter-css>=0.25). That error means "your pip is on an old Python", not a missing dependency. Use pipx/uv, or brew install python@3.12 first.
• OpenAI API key (get one here)

Claude Code plugin (two commands)

/plugin marketplace add curiohunter/hybrid-search-mcp
/plugin install memory-layer@curiohunter

On the next session start the plugin provisions its own Python venv in the background (~1–2 min, once), installs the package, and registers the same global surface as the pip path below — MCP server, hooks, and skills. Restart Claude Code when it reports done, add your OpenAI key (echo "OPENAI_API_KEY=sk-..." >> ~/.env.local), and you're set. Don't combine with the pip setup below — pick one.

pip (three commands)

pipx install memory-layer-mcp                # PyPI name; the CLI is `hybrid-search-mcp`

echo "OPENAI_API_KEY=sk-..." >> ~/.env.local # once per machine — shared by all projects

cd your-project/ && hybrid-search-mcp setup  # once per project (Claude Code)
hybrid-search-mcp setup --codex              # add this if you also use Codex

pip install memory-layer-mcp works too — but Homebrew/system Pythons reject it with externally-managed-environment, so pipx (brew install pipx) is the reliable default for a CLI tool like this.

Why two names? hybrid-search-mcp on PyPI belongs to an unrelated project, so the distribution is published as memory-layer-mcp — which also happens to be the more accurate name. Both hybrid-search-mcp and memory-layer-mcp work as the CLI command.

setup wires everything in one shot: MCP server registration, Claude Code hooks, the /search · /maintain skills, this project's memory hooks, Codex hooks, a CLAUDE.md routing block, and .gitignore entries. Restart Claude Code — the first file you open triggers background indexing, and from then on every git commit re-indexes just the changed files. Nothing else to run.

Adding another project later is one command: cd other-project/ && hybrid-search-mcp setup (global pieces are detected and skipped; each project gets its own isolated index, wiki, and Q&A memory).

CLI-only (no Claude Code)

The search engine works standalone:

cd your-project/
hybrid-search-mcp index .                       # ~165s for 1,776 files, ~$0.04
hybrid-search-mcp search "authentication flow"

From source (contributors)

git clone https://github.com/curiohunter/hybrid-search-mcp.git
cd hybrid-search-mcp
python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/python -m pytest tests/ -q

---

CLI Usage

# Index a project
hybrid-search-mcp index .                    # current directory
hybrid-search-mcp index /path/to/project     # specific path
hybrid-search-mcp index . --force            # full re-index

# Search
hybrid-search-mcp search "login handler"
hybrid-search-mcp search "인증 로직"                     # Korean works
hybrid-search-mcp search "handleSubmit" --node-types function
hybrid-search-mcp search "migration" --file-pattern "*.sql"
hybrid-search-mcp search "auth" --exclude-pattern "docs/*"  # drop doc noise
hybrid-search-mcp search "auth" --json                   # JSON output
hybrid-search-mcp search "query" --limit 20

# Graph exploration
hybrid-search-mcp god-nodes --cwd .          # top-N authority chunks
hybrid-search-mcp annotate-wiki --cwd .      # inject god-nodes Top-N into wiki/index.md (idempotent)
hybrid-search-mcp shortest-path <a> <b>      # call-graph path between two symbols
hybrid-search-mcp subgraph <symbol>          # N-hop forward+reverse call graph

# Memory Layer — persistent Q&A log (ON by default; =0 disables)
hybrid-search-mcp qa-list --cwd .            # recent queries, newest first
hybrid-search-mcp qa-list --all              # across every registered project
hybrid-search-mcp qa-show <id-or-hash>       # full entry (accepts hash prefix ≥4)
hybrid-search-mcp qa-grep "authority"        # frontmatter + body match
hybrid-search-mcp qa-stats --cwd .           # total / by type / by month
hybrid-search-mcp qa-prune --older-than 90d  # rotation
hybrid-search-mcp qa-prune --before 2026-01-01 --dry-run

# Status & maintenance
hybrid-search-mcp status                     # show indexed projects + hooks + skills
hybrid-search-mcp reindex --git-delta --cwd . # delta reindex (changed files only)
hybrid-search-mcp stale --cwd .              # check stale wiki pages

Query auto-classification

The search engine automatically adjusts BM25/vector weights based on query type:

Query	Type	BM25 weight
handleLogin	Exact symbol	0.8 (keyword-heavy)
로그인 처리	Korean NL	0.15 (semantic-heavy)
auth middleware	English NL	0.4 (balanced)

---

Claude Code Integration (Optional)

If you use Claude Code, hybrid-search-mcp becomes an MCP tool with auto-indexing.

Setup

hybrid-search-mcp setup

This registers:

• MCP server in ~/.claude.json
• Auto-index hook (indexes new projects on first file read)
• Stale wiki warning hook
• Wiki gap notification hook
• Route reminder hook (nudges Claude to check wiki before Grep/Glob when .hybrid-search/wiki/index.md exists)

Restart Claude Code after setup.

Codex Integration (Optional)

Codex can share the same project memory layer with Claude Code. One command inside a project:

hybrid-search-mcp setup --codex --cwd .

This writes the .codex-plugin/plugin.json manifest and project .codex/hooks.json, registers the MCP server and [features].hooks in your user-level ~/.codex/config.toml (codex 0.145+ reads MCP config from there, not from the project), adds an AGENTS.md routing note, and runs a config/handler smoke test including a live codex mcp list probe. (install-codex-hook remains as a deprecated alias of the install step, without the smoke test.)

The Codex path uses UserPromptSubmit for pre-answer memory injection and Stop for completed-turn persistence. Stop writes qa logs tagged trigger: codex_stop_hook and client: codex, so Claude Code and Codex can search each other's saved project memory after reindexing.

Skills

Copy skills from skills/ directory to ~/.claude/skills/:

Skill	When	Frequency
/setup-hybrid-search	First install	Once
/bootstrap-wiki	Project onboarding	Per project
/search	Code/doc search with intent routing	Every time
/save-wiki	Save analysis to wiki	Optional
/maintain	Index/wiki maintenance	Occasionally
/rebuild-index	Recovery when index is corrupted or out of sync	Rare

Automation

Trigger	Action	User action
Commit	Git delta reindex + affected wiki refresh	None
Branch checkout	Reindex switched branch	None
Before Grep/Glob	Wiki-first reminder injected into context	None
Before Edit/Write	STALE.md warning	Update wiki
After Read/Edit/Write	Undocumented module alert (wiki-gaps.txt)	Add wiki

---

How It Works

Search strategy — intent-based routing

Query type	Primary	Fallback	Example
Structure/relations	Wiki	hybrid_search	"Who calls this function?"
Feature exploration	hybrid_search	Wiki	"Explain the billing feature"
Exact lookup	Grep	Read	"Where is handleSubmit?"
Design/context	hybrid_search	Wiki	"Why is it designed this way?"
Schema/DB	hybrid_search	Grep	"problems table history"

Benchmark (1,776 files)

Metric	hybrid+Wiki	Grep+Read
Tool calls	2-3	10-15
Time	~3s	20-30s
Accuracy	90%+	Noisy
Token usage	Low	High

Real-world benchmark (valuein_homepage, 1,307 files, 2026-04-22)

20 gold queries across 4 categories (structure / exploration / precision / rationale). Full phase history: Phase 3 (M9 two-pass callgraph + M10 rationale) → Phase 5 (subsystem-first retrieval: module discovery + deterministic card synthesis + module-first injection with Korean↔English alias map).

Headline ratio vs a naive token-bag grep baseline:

Metric (top-10, n=20)	hybrid (Phase 5)	grep (token-bag)	ratio
recall@10 (mean)	0.77	0.37	2.1×
any-hit rate	0.90	0.45	2.0×
primary top-5	0.65	0.35	1.9×
read_count_estimate	4.60	7.40	1.6× (fewer reads)
precision + rationale recall	1.00	0.40	2.5×

Phase-over-phase delta (hybrid track only, showing where subsystem-first helped):

Category	recall@10 Phase 4	recall@10 Phase 5	delta
structure	0.22	0.41	+0.19 (~2×)
exploration	0.47	0.67	+0.20
precision	1.00	1.00	—
rationale	1.00	1.00	—

Honest trade-off: overall read_count_estimate worsened from 3.65 → 4.60 because module cards take rank 1, pushing the top chunk to rank 2 for queries whose real answer is a single doc. Precision + rationale recall preserved.

Compounding benchmark (2026-04-22)

The question this benchmark answers: does the Memory Layer actually make the system smarter as it's used — or is "질문할수록 똑똑해진다" a marketing claim with no numbers behind it?

Methodology (inspired by LongMemEval / LoCoMo session-separated recall):

1. Cold: move .hybrid-search/qa/ aside, reindex with empty memory, run 20 queries.
2. Plant: run 20 Q1a "planter" queries — each logs one Q&A markdown.
3. Warm: reindex so the new qa files become searchable chunks, run 20 queries against warm memory.
4. Score strict gold retrieval (baseline regression guard) separately from memory surface rate and answer_found (gold ∪ memory).

Two tracks:

• Track A — identity: Q1b = Q1a (user asks the same thing again). Upper bound for memory recall.
• Track B — paraphrase: Q1b keeps the principal noun phrases of Q1a but rewords the rest (realistic return-to-same-subsystem scenario). Subset non-leaky excludes pairs where Q1b trivially contains the primary_target filename.

Results (benchmarks/run_compounding_bench.py on valuein_homepage, 20 pairs, 783/783 tests green):

Metric	Cold	Warm	Δ
Track A — identity (user repeats the question)
answer_found (gold or memory in top-10)	80.0%	95.0%	+15.0pp
memory surface rate	0.0%	80.0%	+80.0pp
gold recall@10 (regression guard)	0.656	0.639	−0.017
Track B — paraphrase (reworded follow-up)
answer_found	75.0%	95.0%	+20.0pp
memory surface rate	0.0%	65.0%	+65.0pp
gold recall@10 (regression guard)	0.613	0.596	−0.017
Track B non-leaky (15 pairs, filename not in Q1b)
answer_found	73.3%	100.0%	+26.7pp
memory surface rate	0.0%	66.7%	+66.7pp

What this means:

• Repeated questions: 4 out of 5 times, you see your own past answer without re-searching the codebase. No more retyping "wait, what did I ask last time?"
• Reworded follow-ups: 2 out of 3 times, memory still surfaces — the principal noun phrase is enough signal for the memory boost.
• No regression: strict retrieval on gold code/docs drops 1.7pp (within measurement noise). Memory expands answers, doesn't replace search.
• Honest non-leaky subset: every one of the 15 pairs that don't contain a primary_target filename in Q1b found an answer in top-10 after compounding. 73% → 100%.

Run it yourself:

python benchmarks/run_compounding_bench.py
# → benchmarks/compounding_report_YYYY-MM-DD.md

The script backs up and restores your existing qa directory around the experiment, so it's safe to run against a project you actively use.

Memory bench v2 (2026-07-11)

The compounding bench measures recall. Memory products fail in three other ways that LongMemEval scores and most tools never report: serving a stale fact after it changed, hallucinating confidence on topics the project never had, and burning tokens per answer. Bench v2 (benchmarks/run_memory_bench_v2.py) measures all three on the same production codebase — conflicting Q&A pairs planted 90 days apart, absent-topic probes verified by grep against the corpus, and the exact MCP wire payload counted with o200k.

Scope disclaimer: one production codebase; the update/adversarial cases are synthetic and hand-authored. Read the rates as case counts, not population estimates.

Axis	Metric	Result
Knowledge-update	newer answer shown before the stale one	6/6
	stale answer shown while newer missed (worst case)	0/6
Adversarial recency	old exact-topic answer stays above a fresh adjacent-topic one	3/3 (both retrieved in 1/3; exact-first given both: 1/1)
Tokens per answer	MCP payload, detail=compact (default)	≈ 3.5 k
	detail=full	≈ 4.6 k
Latency	end-to-end search p50 / p95 (incl. confidence pipeline)	567 ms / 863 ms

Abstention — the full confidence distribution, because headline rates alone can be gamed (an all-mixed classifier scores 0 % on both error rates while being useless):

probes	strong	mixed	weak
verified-absent (n=9)	0	1	8
verified-present (n=4)	0	4	0

Honest notes: knowledge-update started at 50 % — the first run of this bench is what exposed that qa recency read filesystem mtime (which lies after a git clone) and that plain topical questions never surfaced memory at all. Supersession is topic-aware, not global: recency reorders only Q&As whose question and answer text overlap, so it can miss weakly-worded update pairs, and the adversarial track exists to prove the reverse failure (fresh-but-adjacent displacing old-but-exact) stays at zero. The absent probe that reads mixed asks about a combination the project lacks ("구독 결제 갱신") while every individual word exists in the source — capping that to weak would require phrase-level reasoning. And yes, the present controls currently top out at mixed, not strong — the strong bar is strict by design, and after the trust-contract change (only verified/accepted memories may anchor strong) its coverage is materially lower on qa-heavy corpora. We do not currently claim "calibrated" confidence: that word is gated on a calibration report (precision AND coverage) at the release head, which has not been re-run since the contract change. The trade-off is visible above instead of averaged away.

python benchmarks/run_memory_bench_v2.py
# → benchmarks/memory_bench_v2_YYYY-MM-DD.md

Language generality — dev sets vs untouched holdout (2026-07-13)

The numbers above come from a Korean development set (valuein). The first English holdout (encode/httpx) caught the supersession topic matcher over-fit to Korean token statistics — knowledge-update collapsed to 1/6 — and that diagnosis burned httpx into a second dev set. After the language-general rewrite (qa_topics.py: English stemming, identifier preservation, weighted overlap, complete-link grouping, calibrated on benchmarks/topic_gold_set.json), a fresh untouched holdout was run once on BurntSushi/ripgrep (different org, different domain, cases authored after freeze, published as-is):

set	role	knowledge-update	adversarial	abstention false-strong
valuein (KO)	dev	6/6	3/3	0/9
httpx (EN)	burned dev	1/6 → 6/6	1/3 → 2/3	0/9
ripgrep (EN)	holdout	synthetic 6/6 (incl. 2 KO→EN probes) · CHANGELOG-derived planted 3/5	2/3	0/9 (incl. 2 KO probes)

A precision note on the KO→EN probes: they prove the end-to-end flow — a Korean question retrieving English Q&A memories whose English old/new pair then groups and supersedes. Grouping a Korean Q&A with an English Q&A is a different capability, exercised by the mixed-language slice of topic_gold_set.json (7/8, two pure cross-language pairs disclosed as known limitations).

Honest notes from the holdout (memory_bench_v2_ripgrep_holdout_*.md): one real stale-fact failure remains (R1 — probe wording matches the obsolete answer verbatim and the correction gets crowded out of top-10 by real corpus hits: a qa-lane exposure gap, not grouping); Korean probes over English memories can miss retrieval entirely (ADV3). Both are documented follow-ups, untouched after the holdout run by rule.

Memory integrity (v0.4.0) — consolidation beyond FIFO

Auto-prune (v0.2.0) keeps the disk bounded; orphan cleanup (v0.3.0) keeps wiki honest. Neither touches the content quality of qa_log over time. v0.4.0 adds three deterministic passes that run at the end of every reindex:

1. Staleness — qa files whose ## Top results references are all gone from the index (typical after a refactor + rename, or a .gitignore addition that drops a tree from indexing) are moved to archive.
2. Semantic dedup — every pair of qa_log chunks is compared on cosine similarity using the vectors already in the HNSW index (no re-embedding, no LLM). Pairs at or above memory.integrity.dedup_threshold (default 0.90) cluster via union-find; the newest of each cluster is kept, rest are archived.
3. Archive TTL — everything archived (by auto-prune, dedup, or staleness) lives in .hybrid-search/qa-archive/YYYY/MM/ for 30 days, then permanently unlinks. qa-restore <id> brings a regretted prune back.

[memory.integrity]
auto_prune = true              # top-level [memory] — unchanged from v0.2.0
enabled = true                 # new — v0.4.0 pass toggle
dedup_threshold = 0.90         # cosine similarity floor for near-duplicates
archive_ttl_days = 30

Run on demand:

hybrid-search-mcp integrity --cwd .                    # defaults
hybrid-search-mcp integrity --cwd . --dedup-threshold 0.85   # more aggressive
hybrid-search-mcp qa-restore abc12345                  # ungarbage-can
hybrid-search-mcp qa-stats --cwd .                     # active/archived counters

Retention — Memory doesn't balloon your disk

.hybrid-search/qa/ grows with every query. The reindex hook applies journald-style two-ceiling retention automatically — whichever ceiling binds first prunes:

[memory]
auto_prune = true
retention_days = 90       # delete anything older
max_files = 2000          # keep at most this many newest
require_first_run_confirm = true   # dry-run on first activation

First auto-prune on a project is a dry-run that reports what would be deleted. Activate with:

hybrid-search-mcp qa-prune --older-than 90d --confirm-first-run

After that, every reindex applies the policy silently. Opt out with auto_prune = false in ~/.hybrid-search/config.toml.

Automatic memory consultation — PreToolUse + SessionStart hooks

By default Claude Code users reach for Grep/Read before remembering an MCP exists. A one-liner install fixes that:

hybrid-search-mcp install-memory-hook --cwd your-project/

This merges two hooks into .claude/settings.local.json (non-destructive — existing hooks are preserved):

• SessionStart: at the start of every session, Claude sees a summary of the 20 most-recent Q&A topics in this project.
• PreToolUse (Grep|Read): before every Grep/Read, a quick qa_log lookup runs — if past Q&A match the pattern, Claude is reminded to prefer mcp__hybrid-search__hybrid_search.

Both are silent when they have nothing to say (no past Q&A matches, noise patterns like . or short strings) — no spam in the context window. Output capped at 800 chars per injection. Inspired by the Graphify pattern (71× fewer tokens reported in real sessions by combining SessionStart + PreToolUse).

Codex memory consultation — UserPromptSubmit + Stop hooks

Codex hooks are installed separately from Claude Code hooks:

hybrid-search-mcp setup --codex --cwd your-project/

Codex project hooks live in .codex/hooks.json; MCP registration and the hooks feature flag live in your user-level ~/.codex/config.toml (codex 0.145+ ignores project-level MCP config). Project-local Codex hooks only run after Codex trusts the project config layer, so use hybrid-search-mcp status --cwd your-project/ and a smoke test before relying on a new install.

Reports:

• Phase 5 full write-up with per-query detail + honest failure modes: benchmarks/valuein_report_v2_2026-04-22.md
• Phase 4 baseline for comparison: benchmarks/valuein_report_2026-04-22.md

Memory Layer

Persist hybrid_search responses as markdown and use them as first-class search targets. Four axes. Write and self-reference are on by default (that is what makes the compounding loop work out of the box) — each is independently opt-out:

export HYBRID_SEARCH_QA_LOG=0      # write:    stop persisting responses
export HYBRID_SEARCH_INDEX_QA=0    # self-ref: stop indexing past qa logs

1. Write

Each response lands at <project>/.hybrid-search/qa/YYYY/MM/DD-HHMMSS-<hash>.md with YAML frontmatter (query, query_type, effective BM25 weight, timestamp)

• top-10 result snippets. A daemon thread does the I/O so the search hot path is not touched. Default on (see Privacy & Data).

2. Read (human)

qa-list [--all] [--since 2026-04-01] [--limit 20] [--json]
qa-show <id | file-stem | hash-prefix≥4>
qa-grep <term> [--case-sensitive]
qa-stats

qa-list --all aggregates across every registered project in newest-first order and prefixes each line with the project name.

3. Self-reference (AI)

Enabled by default (HYBRID_SEARCH_INDEX_QA=0 disables): the scanner walks into .hybrid-search/qa/ (overriding the .gitignore entry that setup writes). Each log becomes a single whole-file chunk tagged node_type="qa_log", so future hybrid_search queries surface past conversations alongside code. Search JSON preserves node_type — clients can filter or re-rank qa hits separately.

4. Rotation

qa-prune --older-than 30d      # or --before 2026-01-01
qa-prune --older-than 90d --dry-run --verbose

Durations accept d / h / w / m (months are 30d approximations). Empty YYYY/MM directories are rmdir'd after a real prune; the qa/ root is preserved as an anchor.

Caveats

• The daemon write is reliable under the long-running MCP server. Short-lived CLI invocations may race with process exit — use the MCP server for production writes.
• qa logs may contain user content — keep both toggles off if you do not want those to leak into general-purpose searches.

Tunables

~/.hybrid-search/config.toml:

[search]
authority_alpha = 0.3  # god-node boost weight. 0.0 disables.
                       # Validated on n=60 (NDCG +0.061, P=1.00).
                       # Externally-weighted workloads may prefer 0.5.

---

Roadmap

Ordered by how often people ask:

1. OpenAI-compatible embedding endpoint — point [embedding] at any base_url (Ollama, vLLM, LM Studio, Voyage/Jina proxies). Unblocks local-only and no-OpenAI users without us shipping a model.
2. Local ONNX embedding backend — small multilingual model for privacy-hard environments (recall trade-off documented honestly).
3. Per-folder embedding policy — mark subtrees (patient data, contracts) as never-embed at config level, on top of the existing .hybrid-search-ignore.

---

Tech Stack

Component	Stack
Embedding	OpenAI text-embedding-3-small
BM25	tantivy-py (Rust)
Vector DB	USearch HNSW (C++)
AST parsing	tree-sitter (C), 14 languages
Storage	SQLite WAL

Supported languages: TypeScript, JavaScript, Python, Rust, Go, Ruby, Java, C, C++, Swift, Kotlin, CSS, HTML, SQL

---

Performance

Operation	Time	Cost
First index (1,776 files)	~165s	~$0.04 (OpenAI embed)
Git delta (post-commit)	~2s	Minimal
Search (direct CLI / MCP call)	<2s	Free
Per-prompt pre-fetch hook (UserPromptSubmit)	~400 ms	Free
grep baseline (for context)	~50 ms	Free

The ~400 ms pre-fetch is the price you pay on every user prompt — that's what lets memory and routing context arrive before the agent picks a tool. For grep-shaped lookups (exact symbol, file path, error string) the router detects this and the hook stays lightweight; for exploratory questions the overhead pays for itself in 1.5–2× time saved downstream (valuein field report v2).

If you want a hard off-switch: export HYBRID_SEARCH_ROUTER=0 disables the pre-fetch entirely.

---

Data locations

~/.hybrid-search/                        # Global
├── config.toml
└── projects/{hash}/store.db

<project>/.hybrid-search/                # Per project
├── wiki/
│   ├── index.md                         # god-nodes Top-N auto-injected via annotate-wiki
│   ├── STALE.md
│   └── {module}.md
├── qa/YYYY/MM/                          # Q&A log (on by default, HYBRID_SEARCH_QA_LOG=0 disables)
└── wiki-gaps.txt

---

CLI Reference

Command	Description
index <path>	Index a project
search <query>	Hybrid search (--file-pattern, --exclude-pattern, --node-types, --json)
serve	Start MCP server (for Claude Code)
setup	Register MCP server + PreToolUse hooks in Claude Code
status	Show indexed projects, hook health, skill install state
reindex --cwd .	Delta reindex
reindex --force --cwd .	Full reindex
stale --cwd .	Check stale wiki pages
viewer --cwd . [--open]	Local memory dashboard (.hybrid-search/viewer.html)
install-hook --cwd .	Install post-commit + post-checkout hooks + .gitignore entries
setup --codex --cwd .	Codex: plugin manifest + hooks + user-level MCP config + smoke test
annotate-wiki --cwd .	Inject god-nodes Top-N into wiki/index.md (idempotent)
god-nodes --cwd .	Top-N authority chunks by call-graph in-degree
shortest-path <a> <b>	Shortest call-graph path between two symbols
subgraph <symbol>	N-hop forward + reverse call graph
synthesize-wiki --cwd .	LLM synthesis for wiki pages
qa-list [--all]	Recent qa logs (Memory Layer); --all aggregates across projects
qa-show <id>	Full qa log by id / stem / hash prefix (≥4 chars)
qa-grep <term>	Substring search over frontmatter + body
qa-stats	Totals by query_type and month
qa-prune --older-than 30d	Delete logs older than a duration or --before <ISO>
remove-project <name>	Unregister a project

---

Troubleshooting

Problem	Solution
OPENAI_API_KEY not found	Set env var or create ~/.env.local
externally-managed-environment on pip install	Homebrew/system Python blocks global pip — use pipx install memory-layer-mcp
"hook error (non-blocking)" on every Read/Edit	Pre-0.5.1 hooks exited non-zero when idle — upgrade, then re-run hybrid-search-mcp setup
Results from wrong project	Use --cwd or --project to scope
Too few results	hybrid-search-mcp index . --force
Rate limit errors	Auto-retry with 0.2s batch interval
Hooks not working	hybrid-search-mcp setup (re-run)
Docs dominate search	--exclude-pattern "docs/*" or "plan/*"
qa log not written from CLI	Expected — async daemon races short-lived CLI exit. Writes via the long-running MCP server are reliable.
qa logs not surfacing in search	Ensure HYBRID_SEARCH_INDEX_QA isn't set to 0, then re-run reindex --force --cwd .

---

License

MIT