Skip to main content
Glama
agy-brain-memory

agy-brain-memory

by orakulzero

agy-brain-memory

Antigravity Brain Cross-Conversation FTS5 Memory, Subagent Lineage & Context Compaction Engine (v1.0.0).

CI License: MIT

Overview

Google Antigravity stores session transcripts (transcript.jsonl) and Markdown artifacts across UUID directories in ~/.gemini/antigravity/brain/<conv-id>. Without an index, searching past sessions or reconstructing parent-to-child invoke_subagent trees requires brute-force linear scans across hundreds of megabytes of JSONL files.

agy-brain-memory solves this with zero third-party runtime dependencies (100% Python 3.10+ standard library sqlite3 WAL + FTS5):

  1. Incremental SQLite WAL + FTS5 Brain Indexer (agy_brain_memory.indexer):

    • Tracks mtime_ns and file_size per conversation transcript so unchanged sessions are skipped in sub-millisecond time.

    • Tolerates malformed or mid-write truncated JSONL lines without crashing.

    • Indexes conversation steps, reasoning summaries, tool calls, and top-level Markdown artifacts (task.md, implementation_plan.md, walkthrough.md).

  2. FTS5 BM25 Cross-Conversation Search & Subagent Lineage (agy_brain_memory.search):

    • Sanitizes arbitrary punctuation/operator queries (:, -, (), [], *) into safe FTS5 token queries.

    • Reconstructs parent-to-child invoke_subagent trees (parent_conv_id -> child_conv_id), subagent roles, prompts, and child final outcomes.

  3. Brain Health Analyzer & Compact Session Digest (agy_brain_memory.compactor):

    • Identifies empty/orphaned directories, BigInt serialization failures, and stalled IN_PROGRESS sessions.

    • Exports token-budgeted Markdown handoff digests (export_session_digest) with conversation://<conv-id> links.

  4. Turnkey Level-2 (L2) Antigravity Plugin & Stdio MCP Server (agy_brain_memory.plugin_installer, agy_brain_memory.mcp_server):

    • Exposes 5 MCP tools (agy_brain_sync, agy_brain_search, agy_brain_subagent_lineage, agy_brain_session_digest, agy_brain_health_report) with composite tool names guaranteed <= 64 characters.

Related MCP server: KusiScribe Core

Quickstart

# 1. Sync ~/.gemini/antigravity/brain into SQLite WAL + FTS5
python -m agy_brain_memory.cli sync

# 2. Search across all conversations and artifacts
python -m agy_brain_memory.cli search --query "HookMultiplexer CRIT-01"

# 3. Inspect subagent parent-child call graph
python -m agy_brain_memory.cli lineage --conv-id <conversation-uuid>

# 4. Export a compact session handoff digest (<= 2000 chars)
python -m agy_brain_memory.cli digest --conv-id <conversation-uuid> --max-chars 2000

# 5. Run brain storage health diagnostics
python -m agy_brain_memory.cli health

# 6. Install turnkey Level-2 native Antigravity plugin
python -m agy_brain_memory.cli install-plugin

# 7. Run the 4-sandbox benchmark suite
python -m agy_brain_memory.cli benchmark

License

MIT License — Copyright (c) 2026 orakulzero.

Available Tools

5 tools
agy_brain_health_reportB

Analyze Antigravity brain storage health, detecting orphaned folders, BigInt crashes, and stalled sessions.

Inputs: no named parameters.

agy_brain_session_digestA

Export a compact Markdown handoff digest of a conversation within a strict character budget.

Inputs: conv_id (required), max_chars.

agy_brain_subagent_lineageA

Reconstruct parent-to-child subagent call-graph trees and child outcomes for a conversation.

Inputs: conv_id, role_query.

agy_brain_syncB

Incrementally index Antigravity brain transcripts and Markdown artifacts into SQLite WAL + FTS5.

Inputs: force.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool occupies a clearly distinct role in the memory lifecycle: indexing (sync), querying (search), call-graph reconstruction (subagent_lineage), export (session_digest), and diagnostics (health_report). There is no verb collision—e.g. only one tool writes to the index and only one reads from it—so misselection risk is very low.

    Naming Consistency4/5

    All five tools share the uniform `agy_brain_` namespace prefix and snake_case formatting, which makes them easy to recognize as a family. The only minor deviation is the mix of verb-first names (`sync`, `search`) with noun-phrase names (`subagent_lineage`, `session_digest`, `health_report`), though each remains self-explanatory.

    Tool Count5/5

    Five tools is a tight, well-scoped surface for an indexed memory store: ingest, query, two analytical views, and one maintenance/diagnostic tool. Nothing appears redundant and no obvious workflow step is missing a tool.

    Completeness4/5

    The surface covers the core lifecycle (sync → search → lineage/digest → health), but there is no explicit delete/prune or reindex/rebuild operation, even though `health_report` detects orphaned folders and stalled sessions—an agent can see problems but cannot act on them. There is also no direct get-by-id fetch of a full conversation, so retrieval relies on search or digest export.

    Maintenance

    ActivityMaintained
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables autonomous multi-agent workflows to capture, preserve, and retrieve immutable transcripts and atomic memory cards through a local MCP server, providing tools for forensic search, topic mapping, and structured fact access without losing context fidelity.
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI coding agents to maintain lossless, queryable long-term project memory through local SQLite storage, bi-temporal history, hybrid retrieval, and Git-backed Markdown mirrors, exposed via MCP tools and resources.
      131 npm
      1
      Apache 2.0
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables agents to persist, search, and consolidate durable cross-session memory through a dual-vault SQLite/FTS5 store with hybrid reranking, retrieval gating, temporal relations, and Markdown wiki projection.
      MIT