engram-rs

Memory engine for AI agents. Two axes: time (three-layer decay & promotion) and space (self-organizing topic tree). Important memories get promoted, noise fades, related knowledge clusters automatically.

Most agent memory is a flat store — dump everything in, keyword search to get it back. No forgetting, no organization, no lifecycle. engram-rs adds the part that makes memory actually useful: the ability to forget what doesn't matter and surface what does.

Single Rust binary, one SQLite file, zero external dependencies. No Python, no Redis, no vector DB — curl | bash and it runs. ~10 MB binary, ~100 MB RSS, single-digit ms search latency.

Quick Start

# Install (interactive — will prompt for embedding provider config)
curl -fsSL https://raw.githubusercontent.com/kael-bit/engram-rs/main/install.sh | bash

# Store a memory
curl -X POST http://localhost:3917/memories \
  -d '{"content": "Always run tests before deploying", "tags": ["deploy"]}'

# Recall by meaning
curl -X POST http://localhost:3917/recall \
  -d '{"query": "deployment checklist"}'

# Restore full context (session start)
curl http://localhost:3917/resume

What It Does

Three-Layer Lifecycle

Inspired by the Atkinson–Shiffrin memory model, memories are managed across three layers by importance:

Buffer (short-term) → Working (active knowledge) → Core (long-term identity)
      ↓                       ↓                           ↑
   eviction              importance decay           LLM quality gate

• Buffer: Entry point for all new memories. Temporary staging — evicted when below threshold
• Working: Promoted via consolidation. Never deleted, importance decays at different rates by kind
• Core: Promoted through LLM quality gate. Never deleted

LLM Quality Gate

Promotion isn't rule-based guesswork — an LLM evaluates each memory in context and decides whether it genuinely warrants long-term retention.

Buffer → [LLM gate: "Is this a decision, lesson, or preference?"] → Working
Working → [sustained access + LLM gate] → Core

Automatic Decay

Decay is activity-driven — it only fires during active consolidation cycles, not wall-clock time. If the system is idle, memories stay intact.

Exponential decay follows the Ebbinghaus forgetting curve — fast at first, then long-tail. Memories never fully vanish (floor = 0.01), remaining retrievable under precise queries. When a memory is recalled, it gets an activation boost, strengthening frequently-used knowledge.

Algorithm Visualizations

Chart	What it shows
	Sigmoid score compression. Raw scores are mapped through a sigmoid function, approaching 1.0 asymptotically. High-relevance results remain distinguishable instead of being crushed into the same value.
	Ebbinghaus forgetting curve. Exponential decay with kind-differentiated rates — episodic memories fade fastest, procedural slowest. Floor at 0.01 means memories never fully vanish; they remain retrievable under precise queries.
	Kind × layer weight bias. Additive biases adjust memory weight by type and layer. Procedural+core memories rank highest, episodic+buffer lowest — but the spread stays bounded so no single combination dominates.
	Reinforcement signals. Repetition and access bonuses follow logarithmic saturation. Early interactions matter most; later ones contribute diminishing returns, discriminating between "used occasionally" and "used daily".
	Use it or lose it. Left: a memory that's never recalled decays into the buffer layer. Right: periodic recall triggers activation boosts that keep the memory in the working layer. Dashed line shows the unrecalled trajectory for comparison.

Semantic Dedup & Merge

Two memories saying the same thing in different words? Detected and merged automatically:

"use PostgreSQL for auth" + "auth service runs on Postgres"
→ Merged into one, preserving context from both

Self-Organizing Topic Tree

Vector clustering groups related memories together, LLM names the clusters. No manual tagging required:

Memory Architecture
├── Three-layer lifecycle [4]
├── Embedding pipeline [3]
└── Consolidation logic [5]
Deploy & Ops
├── CI/CD procedures [3]
└── Production incidents [2]
User Preferences [6]

The problem this solves: vector search requires asking the right question. Topic trees let agents browse by subject — scan the directory, drill into the right branch.

Triggers

Tag a memory with trigger:deploy, and the agent can recall all deployment lessons before executing:

curl -X POST http://localhost:3917/memories \
  -d '{"content": "LESSON: always backup DB before migration", "tags": ["trigger:deploy", "lesson"]}'

# Pre-deployment check
curl http://localhost:3917/triggers/deploy

Session Recovery

Agent wakes up, calls GET /resume, gets full context back. No file scanning needed:

=== Core (24) ===
deploy: test → build → stop → start (procedural)
LESSON: never force-push to main
...

=== Recent ===
switched auth to OAuth2
published API docs

=== Topics (Core: 24, Working: 57, Buffer: 7) ===
kb1: "Deploy Procedures" [5]
kb2: "Auth Architecture" [3]
kb3: "Memory Design" [8]
...

Triggers: deploy, git-push, database-migration

Agent reads the directory, finds relevant topics, calls POST /topic to expand on demand.

Search & Retrieval

Semantic embeddings + BM25 keyword search with CJK tokenization (jieba). IDF-weighted scoring — rare terms get boosted, common terms auto-downweighted. No stopword lists to maintain.

# Semantic search
curl -X POST http://localhost:3917/recall \
  -d '{"query": "how do we handle auth", "budget_tokens": 2000}'
# Note: min_score defaults to 0.30. Use "min_score": 0.0 to get all results.

# Topic drill-down
curl -X POST http://localhost:3917/topic \
  -d '{"ids": ["kb3"]}'

Background Maintenance

Fully automatic, activity-driven — no writes means the cycle is skipped:

Consolidation (every 30 minutes)

1. Decay — reduce importance of unaccessed memories
2. Dedup — merge near-identical memories (cosine > 0.78)
3. Triage — LLM categorizes new Buffer memories
4. Gate — LLM batch-evaluates promotion candidates
5. Reconcile — resolve ambiguous similar pairs (results cached)
6. Topic tree rebuild — re-cluster and name

Topic Distillation — when a topic grows too large (10+ memories), overlapping content is condensed into fewer, richer entries.

Namespace Isolation

Single instance, multiple projects. Use X-Namespace to isolate:

curl -X POST http://localhost:3917/memories \
  -H "X-Namespace: my-project" \
  -d '{"content": "API uses OAuth2 bearer tokens"}'

Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/kael-bit/engram-rs/main/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/kael-bit/engram-rs/main/install.ps1 | iex

Docker

docker run -d --name engram-rs \
  -p 3917:3917 \
  -v engram-data:/data \
  -e ENGRAM_LLM_URL=https://api.openai.com/v1 \
  -e ENGRAM_LLM_KEY=sk-... \
  ghcr.io/kael-bit/engram-rs:latest

LLM Configuration

engram-rs works without an LLM — keyword search and rule-based lifecycle still function. Adding an LLM unlocks intelligent features (triage, gate, merge, topic naming):

export ENGRAM_LLM_URL=https://api.openai.com/v1
export ENGRAM_LLM_KEY=sk-...

# Any OpenAI-compatible API (Ollama, LM Studio, etc.)
export ENGRAM_LLM_URL=http://localhost:11434/v1

# Anthropic native
export ENGRAM_LLM_PROVIDER=anthropic
export ENGRAM_LLM_KEY=sk-ant-...

Two-tier model config — strong model for judgment, lightweight for text processing:

ENGRAM_GATE_MODEL=gpt-4o          # quality decisions
ENGRAM_LLM_MODEL=gpt-4o-mini     # text processing

For AI Agents

Add to your agent's system prompt: