◇ OmniMemory

The memory & context layer for coding agents — persistent, branch-aware, git-anchored, fully local. Your AI stops forgetting between chats and stops hallucinating architecture it never verified.

Works with Claude Code, OpenCode, Antigravity, Cursor & Windsurf (any AGENTS.md-aware IDE). Toggle with /omni-memory. Browse everything in a minimalist local dashboard.

Install

pip (recommended — gives you the omni-memory CLI):

pip install omni-memory-agent

Claude Code plugin (session hooks + skill, no pip needed for injection):

/plugin marketplace add SinghAbhinav04/Omni-Memory
/plugin install omni-memory@singhabhinav

The core is zero-dependency (Python stdlib + SQLite) and runs with no API key. On Python ≥3.10 it auto-installs tree-sitter for the exact multi-language code graph. Without it (e.g. the plugin, or Python 3.9) a stdlib backend still graphs Python (ast) and JS/TS (regex) — approximate, but real.

Quick start — two steps people mix up

They do different things:

First run, in order:

pip install omni-memory-agent

omni-memory build      # 1. build memory + DOCS from your repo   ← the content
omni-memory bind       # 2. wire it into your IDE (hooks + AGENTS.md)
omni-memory ui         # 3. browse memory, docs, and the code graph
omni-memory doctor     # anytime: verify the setup is healthy

About build: it reads your codebase with an agent/LLM. Run it inside an AI IDE (so the agent does the analysis) or set a model key (omni-memory key gemini). Without either, it still builds the code graph + heuristic docs — just no AI-written facts.

After that it's automatic: memory injects into every prompt and captures itself when a session ends.

What it does

• Remembers decisions, facts, request/data flows, gotchas — automatically at the end of each session, and on demand.
• Branch-aware — memory is scoped to your git branch; tracks branch creator, timeline, and merge status. Merged branches roll into the base.
• Pull-based (token-light) — memory isn't force-fed into every prompt. It's seeded once at session start and the agent pulls it on demand (omni-memory inject "<q>") as verified ground truth — cite [id]s or admit "not in memory". Kept fresh (session start + after commits) so it's reliable. Prefer per-prompt enforcement? omni-memory inject-mode auto.
• Relevant (context-aware) — a BM25F ranker (symbol/file/prose field weighting) surfaces the few memories that match your prompt, boosts those anchored to code near what you're editing (via the call graph), and lifts memories the agent has actually cited before. No embeddings, no key.
• Self-checking (symbol-level) — omni-memory check builds a tree-sitter code graph and flags a memory ⚠ stale only when its symbol — or a symbol that calls it — actually changed in git, not just because the file was touched. Falls back to file-level when tree-sitter isn't present.
• Clean — an extraction-noise filter keeps aspirational prose, doc boilerplate, and pasted chat/issue-thread chatter out of the store (gc also retro-sweeps junk already captured).
• Chain-complete — a decision and the gotcha/constraint on the same symbol are retrieved together, so a ranked slice never fragments a causal chain.
• Protected memory — omni-memory lock <id> (or remember --lock) pins architecture/identity/rules as constitutional: never decays, never evicted.
• Self-cleaning — memories stranded on abandoned branches (deleted unmerged, or long dormant) and long-stale/uncited memories are auto-quarantined (reversible), so false memory doesn't live forever. Memories the agent keeps citing are shielded. gc --dry-run previews; hard-delete stays human-gated.
• Merge reconciliation — because merges are known from git, when a branch merges its memories are reconciled onto the base: duplicates collapse automatically, and memories that contradict the base on the same symbol are flagged as conflicts (omni-memory conflicts / resolve <id> [--keep|--both]) with a ⚠CONFLICT marker in the injected block. omni-memory history <id> shows the supersession lineage.
• Team sharing (git-native) — omni-memory share writes your memories to a committed per-author shard under .omni-memory/team/. Because each teammate owns a distinct file, it merges with zero conflicts; a git pull syncs the team's memory into your store automatically at session start (idempotent, tagged ↗external, attributed to its author). Blob-SHAs are content hashes, so a teammate's memory re-verifies on your clone.
• Graph + dashboard — omni-memory ui opens a local UI: browsable memory docs, the knowledge graph, and the repo/branch graph.

Commands

# setup
omni-memory build            # build memory + DOCS from the repo (MEMORY.md, api-map, linkup)
omni-memory bind [ide]       # wire an IDE: session hooks + AGENTS.md (auto-detects)
omni-memory ui               # dashboard: overview · memory · docs · code & repo graph
omni-memory doctor           # diagnose setup (git, store, graph, hooks, AGENTS.md, AI)
omni-memory status | on | off | branch-aware

# using memory
omni-memory recall <q>                       # search memory instead of grepping
omni-memory remember "…" [--kind …] [--global]   # add one by hand (--global = every project)
omni-memory forget <id> · used <id>… · restore <id|branch>
omni-memory branches                         # git topology + per-branch memory

# keeping it fresh
omni-memory map              # (re)build the knowledge + tree-sitter code graph
omni-memory check            # re-anchor vs git; flag ⚠ stale memories (symbol-level)
omni-memory gc [--dry-run] [--purge]         # quarantine dead/false memory
omni-memory usage [--max-items N] [--budget C]   # per-prompt token footprint + tune

# merges / conflicts
omni-memory conflicts        # memories that contradict each other after a branch merge
omni-memory resolve <id> [--keep|--both]     # settle a conflict (winner supersedes, or keep both)
omni-memory history <id>     # supersession lineage: what a memory replaced, and what replaced it

# team / sharing / reset
omni-memory share            # write your committed per-author memory shard (git-merge-safe)
omni-memory sync             # pull teammates' shared memory (also automatic at session start)
omni-memory team             # who contributed what to this project's memory
omni-memory export [file] [--global]         # portable snapshot (commit it to share)
omni-memory import [file] [--global]         # load an export (idempotent, ids preserved)
omni-memory flush [--scope all|memory|graph] [-y]    # wipe to rebuild from scratch
omni-memory key <gemini|anthropic|openai>    # store a model key for the AI build pass

bind is the friendly setup command; install [--platform …] is the explicit form.

How it works

CAPTURE (session + git) → STORE (SQLite, branch-tagged) → RANK + INJECT + ENFORCE
       → CHECK (staleness vs git) → VISUALIZE (dashboard)

Capture fires from deterministic harness events (PreCompact + SessionEnd → capture) — never lost to compaction. Memory is seeded at SessionStart and pulled on demand thereafter (inject-mode: session·auto·manual). Local-first, no cloud, no paid data. See PLAN.md for the full architecture and roadmap.

Status

Core is in: store · git provenance · branch-aware scoping · capture/inject/enforce · context-aware ranker (BM25F + code-graph proximity + citation feedback) · symbol-level staleness · noise filter · memory hygiene (abandoned-branch & false-memory auto-quarantine, human-gated purge) · tree-sitter code graph (Python/JS/TS, with a stdlib-ast fallback) · dashboard (knowledge graph + repo/branch graph). tree-sitter installs automatically on Python ≥3.10; on 3.9 the base install still works and graphs Python via ast. Roadmap: Antigravity via MCP, more languages, runtime request-flow capture. See PLAN.md.

License

Proprietary © 2026 Abhinav Singh. All rights reserved. OmniMemory is not open source — you may install and run the official published package, but not copy, modify, redistribute, or claim it as your own. See LICENSE.