Crate gemini_memory_rs 

gemini-memory-rs

A contextual memory engine for Gemini Live voice sessions.

The engine’s organising principle is that context is prepared asynchronously and consumed synchronously. Nothing expensive — model calls, search, repository writes — ever happens on the path between the model asking for memory and the memory arriving. By the time a recall_context tool call lands, the answer is already sitting in state.

user speech ─► input transcription ─► retrieval-state extraction
                                           │
                                           ▼
                                   local BM25 search
                                           │
                                           ▼
                             immutable prepared snapshot ─► Gemini

final transcript ─► observation extraction ─► session ledger
                                                   │
                       ┌───────────────────────────┤
                       ▼                           ▼
                session overlay          post-session reconciliation
                (usable now)                       │
                                                   ▼
                                         canonical OKF markdown

Layout

Each module is the design’s correspondingly-named component:

Module	Responsibility
core	Domain vocabulary, deterministic policy, event log
okf	Canonical Markdown memory records and the repository
bm25	Fielded lexical index, ranking, and search explanation
transcript	Partial/final transcript accumulation and debouncing
retrieval	Retrieval plans, fusion, budgeted context assembly
ingestion	Observation extraction, candidate ledger, session overlay
reconcile	Consolidation, conflict resolution, promotion, commit
runtime	Live-session wiring: state keys, control loop, tools
evals	Fixture-driven quality harness

Getting started

use gemini_memory_rs::prelude::*;

let engine = MemoryEngine::in_memory(UserId::new("usr_72ab"));

// A finalized user turn: evidence in, context out.
let session = engine.begin_session(SessionId::new("ses_01"));
session.observe_final_transcript(TurnId(1), "I am pescatarian").await?;

let snapshot = session.prepare(TurnId(2), "what should we eat tonight").await?;
for fact in snapshot.facts.iter() {
    println!("{}", fact.statement);
}

Modules

bm25

Local lexical retrieval.

core

Domain vocabulary, policy and the event log.

engine

The memory engine facade.

evals

The fixture-driven quality harness.

ingestion

Capturing evidence from a live conversation.

okf

The canonical memory format: human-readable OKF Markdown.

prelude

The types a typical application touches.

reconcile

Turning a conversation’s evidence into durable memory.

retrieval

Preparing memory context before the model asks for it.

runtime

Wiring the memory engine into a Gemini Live session.

transcript

Transcript accumulation, stable-prefix tracking, and speculation gating.