memlife

Memory that degrades gracefully. Not another pile that grows forever.

Current version: 0.5.0

What's new in 0.5.0

• Pluggable vector backends. Swap between JSON, binary, and sqlite-vec backends with a single config change. The new binary backend stores embeddings as bit-packed vectors for ~32x smaller storage, while sqlite-vec provides fast approximate nearest neighbours when the extension is available.
• Public Metrics snapshot. store.metrics() returns a structured snapshot of counts, embedding coverage, reflection aggregates, recall counters, DB metadata, and schema migration health. Exposed via the memlife://stats MCP resource.
• Schema migration status. store.migration_status() checks whether the database has all expected tables and columns, flags missing items, and reports SQLite version and page stats. Surfaces in Metrics so operators can detect a stale schema before it breaks.
• pysqlite3 fallback for sqlite-vec. On interpreters whose stdlib sqlite3 is compiled without extension loading, memlife transparently falls back to pysqlite3 so sqlite-vec can still load.
• Reflection audit trail. Reflection passes are recorded with proposed, kept, and dropped items, plus timing and model metadata. Paginated audit queries help debug what the reflection loop is actually doing.
• Namespace default and validation hardening. The default namespace is now _default, names are validated/normalized before any DB is opened, and the config fails fast on unknown vector backends or unsafe paths.

What

memlife is a four-tier lifecycle memory system for AI agents. Instead of treating memory as a monotonically growing database, every entry has a lifecycle — facts decay, journal entries retire, superseded data is pruned, and nothing accumulates forever.

The four tiers:

• Episodes — raw events (what happened)
• Facts — durable truths (what I know)
• Journal — reflected beliefs (what I believe)
• Decay/Prune — confidence fades, stale entries retire, GC cleans up

Why

Every other memory system accumulates. Facts never expire. Confidence never decays. Stale conventions become unquestioned truths. Recall quality degrades over time.

memlife solves this. Memory should be like human memory — it fades, it gets revised, it gets pruned. Not a database that grows until it breaks.

Install

pip install memlife

With adapters (optional):

pip install memlife[ollama]       # Ollama embedder + chat
pip install memlife[openai]       # OpenAI embedder + chat
pip install memlife[sentence-transformers]  # Local embeddings
pip install memlife[mcp]          # MCP server

Quickstart (30 seconds, zero dependencies)

import asyncio
from memlife import MemoryStore, MemoryConfig, DummyEmbedder

async def main():
    store = MemoryStore(
        config=MemoryConfig(db_path="./memlife.db", embedding_model="dummy"),
        embedder=DummyEmbedder(),
    )

    # Store an episode (something happened)
    store.remember(task="User asked about deployment", outcome="success")

    # Store a fact (durable truth)
    await store.store_fact("User deploys via GitHub Actions", confidence=0.8)

    # Store an entity relationship (fact-like but structured)
    store.store_triple("User", "deploys_via", "GitHub Actions", confidence=0.8)

    # Retrieve relevant memories (unified scoring across all layers)
    context = await store.retrieve("deployment")
    print(context)

    store.close()

asyncio.run(main())

No Ollama, no OpenAI, no API key. The DummyEmbedder uses bag-of-words vectors — similar sentences get positive cosine similarity. The full lifecycle — store, retrieve, decay, GC, and entity graph — works without any LLM. Only structured extraction and reflection need a model.

The Lifecycle

┌───────────┐     reflection      ┌───────────┐
│  EPISODE  │ ──────────────────▶│  JOURNAL  │
│  (event)  │   LLM synthesises   │ (belief)  │
└─────┬─────┘   observations &   └─────┬─────┘
│  extract triples   │
      │                                 │
      │ store_fact() / store_triple()  │ confidence decay
      ▼                                 │ (configurable)
┌───────────┐    recall bumps    ┌─────▼─────┐
│   FACT    │ ◀────────────────  │  RETIRE   │
│  (truth)  │   confidence +0.05 │ (floor)   │
└─────┬─────┘                    └─────┬─────┘
│    │
│    │ entity graph (triples)
│    ▼
┌───────────────┐
│ TRIPLE / GRAPH│
│(subject-pred- │
│ object + prov)│
└───────┬───────┘
      │
      │ revise / supersede             │ GC prunes
      ▼                                ▼
┌───────────┐                   ┌───────────┐
│ SUPERSEDED│  configurable     │  PRUNED   │
│ (replaced)│ ──────────────────▶│ (deleted) │
└───────────┘   retention       └───────────┘

UNIFIED SCORE = relevance × confidence × recency
Applied across ALL layers before every response.

NO-LLM MODE: store + retrieve + decay + GC + entity graph work
without any model. Only reflection and structured extraction need an LLM.

No-LLM Mode

The store, retrieval, decay, GC, entity graph, and embedding versioning all work without any LLM. Only the reflection loop and structured extraction need a model.

from memlife import MemoryStore, MemoryConfig

store = MemoryStore(config=MemoryConfig(db_path="./memlife.db"))
store.remember(task="something happened", outcome="success")

# retrieve() is async — use SyncMemoryStore or asyncio.run():
import asyncio
context = asyncio.run(store.retrieve("something"))
store.close()

With an Embedder

import asyncio
from memlife import MemoryStore, MemoryConfig
from memlife.adapters.ollama import OllamaEmbedder

async def main():
    store = MemoryStore(
        config=MemoryConfig(db_path="./memlife.db", embedding_model="mxbai-embed-large:latest"),
        embedder=OllamaEmbedder(model="mxbai-embed-large:latest"),
    )
    await store.store_fact("User prefers dark mode", confidence=0.9)
    context = await store.retrieve("dark mode")
    store.close()

asyncio.run(main())

Also available: OpenAIEmbedder (pip install memlife[openai]) and STEmbedder for local Sentence Transformers (pip install memlife[sentence-transformers]).

With Reflection

import asyncio
from memlife import MemoryStore, MemoryConfig, Reflector, DummyEmbedder, DummyChat

async def main():
    store = MemoryStore(
        config=MemoryConfig(db_path="./memlife.db", embedding_model="dummy"),
        embedder=DummyEmbedder(),
    )
    reflector = Reflector(
        memory=store,
        model_chat=DummyChat(),
        critic=False,
    )
    result = await reflector.reflect()
    store.close()

asyncio.run(main())

For real LLMs, use an adapter:

from memlife.adapters.ollama import OllamaChat

# Provide your own model name — memlife doesn't ship deployment-specific defaults.
chat = OllamaChat(model="your-model-name")
reflector = Reflector(memory=store, model_chat=chat, agent_name="my-agent")

Sync API

For non-async codebases:

from memlife import SyncMemoryStore, MemoryConfig, DummyEmbedder

store = SyncMemoryStore(
    config=MemoryConfig(db_path="./memlife.db", embedding_model="dummy"),
    embedder=DummyEmbedder(),
)
store.remember(task="hello", outcome="success")
fact_id = store.store_fact("Test fact", confidence=0.7)
context = store.retrieve("test")

MCP Server

Expose memlife to any MCP-compatible agent (Claude Desktop, Cursor, etc.):

memlife-mcp-server --db ./memlife.db --embedder ollama --embedding-model mxbai-embed-large:latest

Claude Desktop config:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Linux: ~/.config/Claude/claude_desktop_config.json

Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "memlife": {
      "command": "memlife-mcp-server",
      "args": ["--db", "/path/to/memlife.db", "--embedder", "ollama", "--embedding-model", "mxbai-embed-large:latest"]
    }
  }
}

Tools exposed:

Resources:

Features

• Four-tier lifecycle: Episode → Fact → Journal → Decay/Prune
• Entity graph: normalized entities, aliases, and temporal triples with provenance
• Graph traversal: BFS entity neighbors exposed via MCP, no external graph DB
• Triple lifecycle: closed triples and orphan entities/aliases are GC'd like everything else
• Confidence decay: facts decay with a configurable halflife; triples inherit the same decay
• Unified scoring: relevance × confidence × recency across all layers
• Confidence ceiling (0.99): facts are never immutable
• GC with configurable retention: superseded facts, episodes, runs, metrics, and closed triples
• Embedding versioning: detect stale vectors when the model changes, backfill automatically
• Episode tool index: search "have I used this tool before?"
• Incremental contradiction detection: O(new × n), not O(n²)
• Reflection loop: LLM synthesises observations, hypotheses, and revisions with a critic gate
• Structured extraction: optional MEMORIA extraction turns reflection output into attributable triples
• JSONL import/export: backup and migration
• MCP server: plug into Claude, Cursor, or any MCP client
• Adapters: Ollama, OpenAI, Sentence Transformers
• Sync wrapper: for non-async codebases
• SQLite-backed: single file, zero external services
• Zero dependencies: works out of the box with DummyEmbedder + DummyChat

Comparison

memlife wins on lifecycle, decay, and zero-dependency quickstart. It doesn't pretend to beat everyone at everything — Mem0 has multi-user, Graphiti has deep graph analytics. If you want memory that degrades gracefully instead of accumulating forever, memlife is the one.

Status

v0.5.0. The API may change before v1.0. Not recommended for production yet.

License

MIT