Every AI coding agent on your machine keeps its own private memory. Memory Bridge reads them all — the conversations you've already had — and gives Hermes one searchable index across every one of them.
curl -fsSL antharmaya.com/memory-bridge/install.sh | bash9
Agents auto-discovered
0
API keys required
<5ms
Per prefetch()
100%
Local index
Claude Code knows your architecture. Codex remembers your preferences. Gemini has the project plan. But each one starts every session from zero — and Hermes, your always-on agent, can't see any of it.
Other memory tools start recording the day you install them. Memory Bridge reads the conversations already sitting on your disk — every session you've had with Claude Code, Codex, Gemini and the rest — and brings them in on day one.
Memory Bridge builds a local graph of your projects, tools, decisions and files — and recalls associatively. Ask “what did I do with Claude Code last week,” or get a live brief on any project: its decisions, outcomes, connected tools, and key files, in one block.
Facts are atomic, dated, and source-linked. Briefings are computed live from those facts — never a stored, mutated summary — so they can't drift, hallucinate, or go stale. Reflection only ever adds; it never overwrites.
Typo-tolerant FTS5 trigram search, associative graph recall, and optional local embeddings — fused with Reciprocal Rank Fusion. Sub-millisecond, no vector database, no cloud. Semantic is a single optional install, never required.
Ships a zero-dependency MCP server, so Claude Desktop, Cursor, Codex — any MCP client — can query the same unified index. One brain, every agent you run.
Deep extraction uses ctx.llm — Hermes' own model — so there's no second key to manage. The index is a local SQLite file that never leaves your machine. No telemetry, no account.
Every architecture call, every trade-off, every "we'll go with X" gets extracted, attributed to the framework that shaped it, and tracked through to its outcome.
You make a call — a framework, an architecture, a trade-off — across any of your agents.
Memory Bridge extracts the decision, attributes the framework that shaped it, and files it with its rationale and confidence.
Later, mark how it actually turned out. Your agents stop repeating decisions that already failed.
$ hermes memory-bridge decisions --unverifiedAuto-discovered on your machine — no hooks, no setup
Auto-discovers every agent history on disk — 9 agents, no hooks, no instrumentation.
Rules-based extraction is free and instant; an optional ctx.llm pass adds deep semantic facts.
SQLite + FTS5 trigram. Sub-millisecond, typo-tolerant, zero external services.
Every turn, relevant memory is injected via hybrid recall (lexical + graph + optional semantic). Or any MCP client queries it directly.
histories → scanner → extractor → SQLite + FTS5 → Hermes, every turn
No external services. No API keys. No telemetry. Python 3.11+ and Hermes 0.16+.
curl -fsSL https://antharmaya.com/memory-bridge/install.sh | bashClones into ~/.hermes/plugins/memory/, installs deps, runs the first scan.
git clone https://github.com/antharmaya/mem-bridge.git \
~/.hermes/plugins/memory/antharmaya-bridgeMemory providers live in ~/.hermes/plugins/memory/<name>/ — the Hermes discovery convention.
hermes memory setupSelect "antharmaya-bridge", or set memory.provider in ~/.hermes/config.yaml.
MIT licensed · Built by Antharmaya Labs · Made in India