CogZ

CogZ

Local-first, code-aware engineering cognition for AI coding agents.

CogZ gives a coding agent persistent memory, contextual retrieval, and continuous cognition about a software repository — all running locally on your machine, no cloud services required.

Works with Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, Devin, and any MCP-compatible agent.

What it looks like

Real output from CogZ running on its own codebase:

$ cogz context --mode task "token budget estimation and context pack compression"

Context pack (mode: task) Query: token budget estimation and context pack compression Search mode: hybrid Sections: 91 Token estimate: 8188

Dropped: 6 sections over token budget

---

1. [rule] New expansion channels: emit early, filter before seen-mark, sort deterministically, never displace directs (relevance: 0.5148)

Conventions proven across the sibling and co-change channels:

1. Emit before the generic expansion loops — candidates emitted later get claimed-and-floored by graph traversal …
2. Apply entity-type/test filters BEFORE seen.insert …

…

2. [rule] cfg-gated code must be typechecked per-target before release (relevance: 0.4690)

Code behind #[cfg(unix)]/cfg(target_os = ...) is invisible to host builds, tests, and clippy — a compile error in a cfg'd branch ships silently until a real target build sees it. The v0.5.0 Windows leg failure is the canonical example.

3. [rule] Degradation must be loud, never silent (relevance: 0.3680)

Every degraded or failed code path must surface a signal …

4. [identity] CogZ (relevance: —)

Project: CogZ

5. [file] assemble.rs (relevance: 0.6993)

//! Context pack assembly — the tiered-push pipeline. //! Tier 0 (baseline: identity + top rules) always ships for task and //! escalation packs …

… 86 more sections …

That's not a text chunk from a vector search. The pack leads with validated rules — one learned from a release failure on this very project — plus the identity baseline and the actual source file, all ranked, traceable, and budgeted.

This repository already contains real dogfooding knowledge — CogZ has been used on its own codebase throughout development. You can clone it, install CogZ, and try the commands above against it directly.

What it does

CogZ maintains a project-specific knowledge layer that connects what an agent learns to the code it is working with.

Memory

CogZ stores three kinds of project knowledge:

• Observations — things an agent has learned or noticed. Raw, unvalidated experience: bugs found, decisions made, patterns noticed.
• Rules — validated knowledge that should influence future work. Coding standards, design decisions, confirmed patterns.
• Knowledge — structured information about the codebase. Architecture explanations, module responsibilities, trade-off rationale.

These are stored as Markdown files with YAML frontmatter, linked to each other and to code entities in the repository. The files are the canonical source of truth — SQLite is a derived index, disposable and rebuildable. Your knowledge is portable, version-controlled, and editable by hand.

Context

Instead of giving an agent everything it knows, CogZ builds scoped context packs for the current situation. A context pack combines relevant rules, observations, knowledge, and code structures — ranked by relevance, traceable through the code graph, and limited by a token budget so the agent gets what matters for the task rather than the entire project history.

Cognition

CogZ periodically consolidates what has been learned: deduplicates entries, detects contradictions, promotes well-supported observations to rules, merges superseded entries, and flags knowledge as stale when the code it references changes.

Quick start

Linux / macOS / Windows (Git Bash):

# Install

curl -fsSL https://raw.githubusercontent.com/balaianu/CogZ/master/install.sh | bash

Initialize in a repo (add --configure auto to wire MCP + hooks for detected agents)

cd ~/your-project cogz init

Index (downloads models on first run, or use --no-download for FTS-only)

cogz index

Verify it's working — entity counts, model status, DB stats

cogz status

Windows (PowerShell):

# Install

irm https://raw.githubusercontent.com/balaianu/CogZ/master/install.ps1 | iex

Initialize in a repo

cd your-project cogz init cogz index

See Getting Started for the mental model and a complete walkthrough.

MCP integration

CogZ runs as a stateless MCP server over stdio. Every tool call specifies which repo it targets via a required repo parameter — no Roots, no session state, no fallbacks.

{

"mcpServers": { "cogz": { "command": "cogz", "args": ["mcp-stdio"] } } }

The server exposes 15 tools: create_entity, update_knowledge, verify_knowledge, reject_entity, query_entities, search, get_context, get_status, list_entities, consolidate, capture_event, get_callers, get_impact, find_orphans, suggest_observations.

See MCP Tools for full parameter reference and example responses. See Agent Setup for per-agent config files, hook formats, and verified capability notes for all six supported agents — or just run cogz configure auto.

Hook integration

Hooks capture lifecycle events and inject context packs into agent sessions. CogZ's binary is the hook handler — no wrapper scripts needed.

{

"hooks": { "SessionStart": [{ "matcher": "", "hooks": [{ "type": "command", "command": "cogz capture-event session_start --hook-json", "timeout": 15 }] }] } }

See Hooks for all 7 event types and per-agent wiring guides.

CLI commands

Normal operation is automatic: hooks fire on lifecycle events, the agent drives CogZ through MCP. The CLI is not needed for day-to-day use — it's available for setup, manual exploration, and automation if you want or need it.

See CLI Reference for all flags and options.

Requirements

Minimum (FTS-only mode)

Works without ONNX Runtime or model downloads. All hooks, FTS search, context packs, consolidation, doctor, and prune are functional. Vector search, embedding-based dedup, and contradiction detection are not available.

Recommended (hybrid search mode)

Full functionality including vector search, semantic dedup, and NLI contradiction detection. Models auto-download on first use and auto-unload after 5 min idle (RAM drops back to ~11 MB). See Evaluations for the full resource consumption profile.

Benchmarks

CogZ ships a reproducible suite (benchmark/) run on pinned public corpora — httpx, cobra, clap, each injected with memory seeds mined from its real git history — plus this repository's own .cogz corpus. Seeded ground truth:

Channel ablations on commit queries: removing graph expansion costs 3–15pt recall@20; FTS-only mode is competitive on commit-intent queries (0.34–0.47 R@20 vs hybrid's 0.31–0.47) — the vector channel's lift concentrates on semantic queries while expansion dilutes under the row cap (tracked as the top ranking issue). Context packs keep 0.49–0.89 expected-entity recall at the default 8K budget. Reruns are byte-identical. Full methodology, per-phase numbers, and the raw artifacts: benchmark/README.md.

What using it buys (measured): in a 14-task agent replay, the seeded-knowledge arm finished ~2x faster than bare (871s vs 1748s average) and completed more runs (14/14 vs 10/14) at equal correctness. Consolidation machinery is precise: dedup precision/recall 1.0, NLI contradiction detection 4/4 with zero false alarms, drift marking exact.

Honest limits: top-5 precision is weak on mixed corpora (P@5 <= 0.20; code entities outrank knowledge at the top of the ranking), commit-intent queries reach 0.31–0.53 recall@20, adjacent-domain negative queries still leak confident hits (silence-gate clean rate 0.47–0.73 across corpora), and at n=14 tasks there is no measurable task-correctness lift yet.

Architecture

• Single Rust binary — no runtime dependencies except optional ONNX models for vector search.
• Files are canonical — all entities are Markdown files. The SQLite DB is a derived index, disposable and rebuildable.
• Code-aware — tree-sitter indexes source code as first-class graph entities. Supported languages: Rust, Python, Go, JavaScript, TypeScript, TSX, Bash.
• Graceful degradation — works without ML models in FTS-only mode.
• Local-first — no cloud, no telemetry, no accounts. The only network access is optional model downloads.

See Architecture for the full system design.

Compatibility

macOS Intel is not supported because Microsoft dropped ONNX Runtime macOS Intel binaries after v1.22. Intel Mac users can run the arm64 binary under Roset