Agent memory you can read, search, and git diff.

memweave is a zero-infrastructure, async-first Python library that gives AI agents persistent, searchable memory — stored as plain Markdown files and indexed by SQLite. No external services. No black-box databases. Every memory is a file you can open, edit, grep, and version-control.

• 📄 Human-readable by design. Memories live in plain .md files on disk. Open them in your editor, inspect them in your terminal, or git diff what your agent learned between runs.
• 🔍 Hybrid search out of the box. Combines BM25 keyword ranking (FTS5) with semantic vector search (sqlite-vec) and merges them — so "PostgreSQL JSONB" finds both exact matches and conceptually related content.
• ⚡ Zero LLM calls on core operations. Writing and searching memories never touches an LLM. Embeddings are cached by content hash — compute once, reuse forever.
• 🌐 Works completely offline. If your embedding API is down, memweave falls back to pure keyword search. It never crashes; it degrades gracefully.
• 💸 Zero server cost, zero setup. The entire memory store is a single SQLite file on disk — no vector database to provision, no cloud service to pay for, no Docker container to manage.
• 🔌 Pluggable at every layer. Swap in a custom search strategy, add a post-processing step, or bring your own embedding provider via a single protocol.
• 📅 Memories age naturally. Recent knowledge ranks above stale context automatically — no manual cleanup, no ever-growing noise. Foundational facts stay exempt.
• 🎯 No redundant results. MMR re-ranking ensures the top results cover different aspects of your query — not the same fact repeated from five slightly different chunks.

• Quickstart Guide
• How it works
• Core concepts
  • Markdown as the source of truth
  • Evergreen vs dated files
  • Agent namespaces & source labels
  • Search pipeline
  • Temporal decay
  • MMR re-ranking
• Usage examples
  • Single agent with persistent memory
  • Multi-agent with shared and isolated namespaces
  • Memory flush
  • Custom search strategy
  • File watcher
  • Inspect memory status
  • List indexed files
• Configuring memweave
  • Embedding providers
• API reference
• Contributing
• License

pip install memweave

Set an embedding provider (or skip to use keyword-only mode):

export OPENAI_API_KEY=sk-...

import asyncio
from pathlib import Path
from memweave import MemWeave, MemoryConfig

async def main():
    async with MemWeave(MemoryConfig(workspace_dir=".")) as mem:
        # Write a memory file, then index it
        memory_file = Path("memory/preferences.md")
        memory_file.parent.mkdir(exist_ok=True)
        memory_file.write_text("The user prefers dark mode and concise answers.")
        await mem.add(memory_file)

        # Search across all memories.
        # min_score=0.0 ensures results surface in a small corpus;
        # in production the default 0.35 threshold filters low-confidence matches.
        results = await mem.search("What is the user preference?", min_score=0.0)
        for r in results:
            print(f"[{r.score:.2f}] {r.snippet}  ← {r.path}:{r.start_line}")

asyncio.run(main())

Memories are plain Markdown files in memory/. Inspect them any time:

cat memory/*.md

Each result includes a relevance score and the exact file and line it came from:

[0.35] The user prefers dark mode and concise answers.  ← memory/preferences.md:1

memweave separates storage from search:

┌──────────────────────────────────────────────────────────────┐
│                 SOURCE OF TRUTH  (Markdown files)            │
│   memory/MEMORY.md        ← evergreen knowledge              │
│   memory/2026-03-21.md    ← daily logs                       │
│   memory/agents/coder/    ← agent-scoped namespace           │
└───────────────────────┬──────────────────────────────────────┘
                        │  chunking → hashing → embedding
┌───────────────────────▼──────────────────────────────────────┐
│                DERIVED INDEX  (SQLite)                       │
│   chunks          — text + metadata                          │
│   chunks_fts      — FTS5 full-text index  (BM25)             │
│   chunks_vec      — sqlite-vec SIMD index (cosine)           │
│   embedding_cache — hash → vector  (skip re-embedding)       │
│   files           — SHA-256 change detection                 │
└───────────────────────┬──────────────────────────────────────┘
                        │  hybrid merge → post-processing
                        ▼
              list[SearchResult]

Write path — await mem.add(path) takes any Markdown file you've written — dated, evergreen, agent-scoped, or session — chunks it, checks the embedding cache (hash lookup), calls the embedding API only on a miss, and inserts into both the FTS5 and vector tables. No LLM involved.

Search path — await mem.search(query) embeds the query, runs vector search and keyword search in parallel, merges scores (0.7 × vector + 0.3 × BM25), applies post-processors (threshold → temporal decay → MMR), and returns ranked results.

The SQLite index is a derived cache — always rebuildable from the Markdown files. This means:

• You can edit memories directly in your editor and re-index with await mem.index().
• git diff memory/ shows exactly what an agent learned between commits.
• Losing the database is not data loss. Losing the files is.

Evergreen files hold foundational facts that should always surface at full score. Dated files accumulate daily learning and fade naturally — recent memories rank higher.

Every file gets a source label derived from its path — the immediate subdirectory under memory/ becomes the label:

Pass source_filter="researcher_agent" to search() to scope results exclusively to that namespace. Only the first path component counts — memory/researcher_agent/sub/x.md has source "researcher_agent", not "sub".

Every mem.search(query) call moves through five fixed stages in order:

                        query
                          │
             ┌────────────┴────────────┐
             │                         │
     FTS5 BM25 (keyword)      sqlite-vec ANN (semantic)
     exact term matching       cosine similarity
             │                         │
             └────────────┬────────────┘
                          │  weighted merge
                          │  score = 0.7 × vector + 0.3 × BM25
                          │
                 ┌────────▼────────┐
                 │ ScoreThreshold  │  drop results below min_score (default 0.35)
                 └────────┬────────┘
                          │
                 ┌────────▼────────┐
                 │ TemporalDecay   │  multiply score by exp(−λ × age_days)
                 │  (opt-in)       │  evergreen files exempt
                 └────────┬────────┘
                          │
                 ┌────────▼────────┐
                 │  MMR Reranker   │  reorder for relevance + diversity
                 │  (opt-in)       │  λ × relevance − (1−λ) × similarity_to_selected
                 └────────┬────────┘
                          │
                 ┌────────▼────────┐
                 │ Custom          │  your own PostProcessor(s)
                 │ processors      │  via mem.register_postprocessor()
                 └────────┬────────┘
                          │
                   list[SearchResult]

Stage 1 — Hybrid merge. Both backends run against the same query. FTS5 BM25 catches exact keyword matches (error codes, config values, proper names). sqlite-vec cosine catches semantically related content even when no keyword overlaps. Scores are normalised and merged: 0.7 × vector_score + 0.3 × bm25_score. Weights are tunable via HybridConfig.

Stage 2 — Score threshold. Drops any result whose merged score is below min_score (default 0.35). Acts as a noise gate — prevents low-confidence matches from entering the post-processing stages. Always active; override per-call with mem.search(query, min_score=0.5).

Stage 3 — Temporal decay (opt-in). Multiplies each result's score by an exponential factor based on the age of its source file. Recent memories rank higher; old ones fade naturally. Evergreen files are exempt and always surface at full score. See Temporal decay below.

Stage 4 — MMR re-ranking (opt-in). Reorders the remaining results to balance relevance against diversity. Prevents the top results from being near-duplicates of each other. See MMR re-ranking below.

Stage 5 — Custom processors. Any processors registered with mem.register_postprocessor() run last, in registration order. Each receives the output of the previous stage and can filter, reorder, or rescore freely.

Agents accumulate knowledge over time — but not all knowledge ages equally. A decision made yesterday should outrank one made six months ago when both are semantically relevant. Without decay, a stale debugging note from last quarter can surface above this morning's architecture decision simply because it embeds well.

Temporal decay solves this by multiplying each result's score by a factor that shrinks the older the source file is. The score is never zeroed out — old memories still surface, they just rank lower than recent ones.

How the formula works:

λ            = ln(2) / half_life_days
multiplier   = exp(−λ × age_days)
decayed_score = original_score × multiplier

At age_days = 0 the multiplier is 1.0 — no change. At age_days = half_life_days it is exactly 0.5. The curve is smooth and continuous, so a file that is two half-lives old scores at 0.25×, three half-lives at 0.125×, and so on.

With the default half_life_days=30:

How age is determined — three file categories:

Evergreen files hold foundational facts — stack choices, hard constraints, permanent preferences — that should always surface at full score regardless of when they were written. Daily logs capture evolving context and fade naturally as new sessions add fresher knowledge.

Enabling temporal decay:

from memweave import MemWeave
from memweave.config import MemoryConfig, QueryConfig, TemporalDecayConfig

config = MemoryConfig(
    query=QueryConfig(
        temporal_decay=TemporalDecayConfig(
            enabled=True,
            half_life_days=30.0,   # score halves every 30 days; tune to your workflow
        ),
    ),
)

async with MemWeave(config) as mem:
    results = await mem.search("database choice")
    # results from last week will rank above results from last quarter
    # results from memory/MEMORY.md are exempt and always surface at full score

Tune half_life_days to your workflow: 7 for fast-moving projects where week-old context is already stale, 90 for research or documentation repositories where knowledge stays relevant for months.

Without diversity control, the top results from a hybrid search are often near-duplicates — multiple chunks from the same file, or different phrasings of the same fact. An agent loading all of them into its context window wastes tokens and misses other relevant but different memories.

MMR (Maximal Marginal Relevance) reorders results after scoring to balance how relevant a result is against how similar it is to results already selected. At each step it picks the candidate that maximises:

MMR score = λ × relevance − (1−λ) × max_similarity_to_already_selected

Similarity is computed as Jaccard overlap between the token sets of the candidate and each already-selected result. This means two chunks that share many of the same words — even from different files — are treated as redundant, and the second one is pushed down in favour of something genuinely different.

The lambda_param dial:

Enabling MMR:

from memweave import MemWeave
from memweave.config import MemoryConfig, QueryConfig, MMRConfig

config = MemoryConfig(
    query=QueryConfig(
        mmr=MMRConfig(
            enabled=True,
            lambda_param=0.7,   # 0 = max diversity, 1 = max relevance
        ),
    ),
)

async with MemWeave(config) as mem:
    results = await mem.search("deployment steps")
    # top results will cover different aspects of deployment
    # rather than returning the same facts from multiple angles

    # override λ per-call without touching the config
    diverse = await mem.search("deployment steps", mmr_lambda=0.3)

MMR runs after temporal decay, so the diversity pass operates on already age-adjusted scores — the reranker sees a realistic picture of which results actually matter before deciding what is redundant.

import asyncio
from pathlib import Path
from memweave import MemWeave, MemoryConfig

async def run_agent_session():
    config = MemoryConfig(workspace_dir="./my_project")

    async with MemWeave(config) as mem:
        # Write memory files, then index them
        memory_dir = Path("my_project/memory")
        memory_dir.mkdir(parents=True, exist_ok=True)

        (memory_dir / "stack.md").write_text("User's preferred stack: FastAPI + PostgreSQL + Redis.")
        (memory_dir / "guidelines.md").write_text("Avoid using global state in this codebase.")

        await mem.index()

        # Retrieve relevant context before responding
        context = await mem.search("database recommendations", min_score=0.0, max_results=2)
        for result in context:
            print(f"  [{result.score:.2f}] {result.snippet}  ({result.path}:{result.start_line})")

asyncio.run(run_agent_session())

Agents share one workspace but write to separate subdirectories under memory/. The subdirectory name becomes the source label — pass source_filter="researcher_agent" to scope a search exclusively to that agent's files.

import asyncio
from pathlib import Path
from memweave import MemWeave, MemoryConfig

async def main():
    # Both agents share the same workspace root
    researcher = MemWeave(MemoryConfig(workspace_dir="./project"))
    writer = MemWeave(MemoryConfig(workspace_dir="./project"))

    async with researcher, writer:
        # Researcher writes space exploration findings to its own namespace
        memory_dir = Path("project/memory/researcher_agent")
        memory_dir.mkdir(parents=True, exist_ok=True)

        (memory_dir / "mars_habitat.md").write_text(
            "Mars surface pressure is ~0.6% of Earth's, requiring fully pressurised habitats. "
            "NASA's MOXIE experiment on Perseverance successfully produced oxygen from CO2 in 2021, "
            "validating in-situ resource utilisation (ISRU) as a viable strategy for long-duration missions."
        )
        (memory_dir / "artemis_mission.md").write_text(
            "Artemis III aims to land the first woman and next man near the lunar south pole. "
            "Permanently shadowed craters there hold water ice deposits confirmed by LCROSS in 2009. "
            "Ice can be electrolysed into hydrogen and oxygen, serving as both breathable air and rocket propellant."
        )
        (memory_dir / "deep_space_propulsion.md").write_text(
            "Ion drives expel charged xenon atoms at ~90,000 km/h, achieving far higher specific impulse "
            "than chemical rockets, though thrust is measured in millinewtons. NASA's Dawn spacecraft used "
            "ion propulsion to orbit both Vesta and Ceres — the first mission to orbit two extraterrestrial bodies."
        )

        await researcher.index()

        # Writer queries the researcher's findings — scoped to the researcher_agent source
        queries = [
            "how do astronauts get oxygen on Mars",
            "water ice on the Moon",
            "spacecraft propulsion beyond chemical rockets",
        ]

        for query in queries:
            print(f"\nQuery: {query!r}")
            results = await writer.search(query, source_filter="researcher_agent", min_score=0.0, max_results=1)
            for r in results:
                print(f"  [{r.score:.2f}] {r.snippet}  ({r.path}:{r.start_line})")

asyncio.run(main())

LLM context windows are finite. When a long conversation is compacted or a session ends, anything not written to memory is lost. flush() solves this by sending the conversation to an LLM with a structured extraction prompt — the model distils durable facts (decisions, preferences, constraints) and discards small talk. The extracted text is appended to the dated memory file (memory/YYYY-MM-DD.md) and immediately re-indexed, so it surfaces in future searches. If the LLM finds nothing worth storing it returns a silent sentinel and flush() returns None — nothing is written.

Requires an LLM API key (configured via FlushConfig.model, default gpt-4o-mini).

import asyncio
from pathlib import Path
from memweave import MemWeave, MemoryConfig

WORKSPACE = Path(__file__).parent / "workspace"

conversation = [
    {"role": "user",      "content": "We just decided to use Valkey instead of Redis for caching."},
    {"role": "assistant", "content": "Got it. I'll note that Valkey is the new caching layer."},
    {"role": "user",      "content": "Also, we're targeting a 5ms p99 latency SLA for the cache."},
]

async def main():
    config = MemoryConfig(workspace_dir=WORKSPACE)

    async with MemWeave(config) as mem:
        # Extract durable facts from the conversation and write to workspace/memory/YYYY-MM-DD.md.
        # Returns the extracted text, or None if there was nothing worth storing.
        extracted = await mem.flush(conversation=conversation)
        if extracted:
            print(f"Stored:\n{extracted}\n")
        else:
            print("Nothing worth storing.\n")

        # Search the indexed knowledge immediately after flush
        results = await mem.search("Valkey caching latency", min_score=0.0)
        print(f"Search results ({len(results)} hits):")
        for r in results:
            print(f"  [{r.score:.3f}] {r.snippet.strip()}")

asyncio.run(main())

The built-in "hybrid", "vector", and "keyword" strategies cover most cases, but sometimes you need ranking logic that none of them support — for example, boosting results from recently modified files, hard-pinning results from a specific file to the top, or implementing a completely different scoring algorithm. A custom strategy gives you direct access to the SQLite database, so you can write any query you like and return results in whatever order you want. memweave applies your results through the same post-processing pipeline (score threshold, MMR, temporal decay) as built-in strategies.

Register a strategy once with mem.register_strategy(name, obj), then activate it per-call via strategy=name.

import asyncio
import aiosqlite
from memweave import MemWeave, MemoryConfig
from memweave.search.strategy import RawSearchRow

class RecencyBoostStrategy:
    async def search(
        self,
        db: aiosqlite.Connection,
        query: str,
        query_vec: list[float] | None,
        model: str,
        limit: int,
        *,
        source_filter: str | None = None,
    ) -> list[RawSearchRow]:
        # Your custom ranking logic here — query `db` directly and return RawSearchRow objects
        ...

async def main():
    async with MemWeave(MemoryConfig(workspace_dir=".")) as mem:
        mem.register_strategy("recency", RecencyBoostStrategy())
        results = await mem.search("recent decisions", strategy="recency")

asyncio.run(main())

When running a long-lived agent, memory files can be edited externally — by another process, a human, or a separate agent writing to the same workspace. Without the watcher, those changes are invisible until the next explicit await mem.index() call. start_watching() launches a background task that monitors the memory/ directory and re-indexes any .md file the moment it changes, so searches always reflect the latest content. Rapid successive writes are debounced (default 1500 ms) to avoid redundant re-indexing. The watcher stops automatically when the context manager exits.

Requires the watchfiles package (pip install memweave[watch]).

import asyncio
from memweave import MemWeave

async def main():
    async with MemWeave() as mem:
        await mem.start_watching()   # starts background task, watches memory/
        # ... run your agent loop
        # any .md file edits are picked up and re-indexed automatically
        # watcher stops automatically on context manager exit

asyncio.run(main())

status() gives a point-in-time snapshot of the store — how many files and chunks are indexed, which search mode is active (hybrid, fts-only, or vector-only), whether there are unindexed changes pending (dirty), and how many embeddings are cached. Useful for health checks, debugging, or surfacing store state in agent logs.

async with MemWeave() as mem:
    status = await mem.status()
    print(f"Files:       {status.files}")
    print(f"Chunks:      {status.chunks}")
    print(f"Search mode: {status.search_mode}")   # hybrid | fts-only | vector-only
    print(f"Dirty:       {status.dirty}")         # unindexed changes pending

files() returns metadata for every file currently tracked in the index — path, size, chunk count, source label, and whether the file is evergreen. Useful when an agent needs to audit what it has access to, detect stale files, or decide which namespace to write to next.

async with MemWeave() as mem:
    for f in await mem.files():
        print(f"{f.path}  ({f.chunks} chunks, evergreen={f.is_evergreen}, source={f.source})")

All configuration is optional — sensible defaults work out of the box. Pass a MemoryConfig to override.

MemoryConfig is a single nested dataclass that groups every tunable knob into focused sub-configs. Each sub-config has its own defaults and can be overridden independently:

• EmbeddingConfig — which model to use for vectorising text, API key, batch size, timeout.
• ChunkingConfig — chunk size and overlap in tokens. Smaller chunks give more precise retrieval; larger chunks give more context per result.
• QueryConfig — default search strategy, max results, score threshold, and the settings for the three built-in post-processors (hybrid weights, MMR, temporal decay).
• CacheConfig — embedding cache toggle and optional LRU eviction cap to bound disk usage.
• SyncConfig — when to auto-reindex (before each search, on file change, or on a periodic interval).
• FlushConfig — the LLM model and system prompt used by flush() for fact extraction.

Every field can also be overridden per-call at search time (e.g. min_score, max_results, strategy) without touching the config.

from memweave import MemWeave
from memweave.config import (
    MemoryConfig, EmbeddingConfig, QueryConfig,
    HybridConfig, MMRConfig, TemporalDecayConfig,
    SyncConfig, FlushConfig,
)

config = MemoryConfig(
    workspace_dir="./memory",            # where .md files live

    embedding=EmbeddingConfig(
        model="text-embedding-3-small",  # any LiteLLM-compatible model
        api_key="sk-...",                # or set via environment variable
        batch_size=100,
    ),

    query=QueryConfig(
        strategy="hybrid",               # "hybrid" | "vector" | "keyword"
        max_results=10,
        min_score=0.35,

        hybrid=HybridConfig(
            vector_weight=0.7,           # weight for semantic similarity
            text_weight=0.3,             # weight for BM25 keyword score
        ),

        mmr=MMRConfig(
            enabled=True,
            lambda_param=0.5,            # 0 = max diversity, 1 = max relevance
        ),

        temporal_decay=TemporalDecayConfig(
            enabled=True,
            half_life_days=30.0,         # score halves every 30 days
        ),
    ),

    sync=SyncConfig(
        on_search=True,                  # sync dirty files before each search
        watch=False,                     # enable file watcher
        watch_debounce_ms=500,
    ),

    flush=FlushConfig(
        enabled=True,
        model="gpt-4o-mini",             # LLM used for fact extraction
    ),
)

async with MemWeave(config) as mem:
    ...

memweave uses LiteLLM under the hood — any LiteLLM-compatible embedding model works with zero code changes:

Ollama (no API key required):

from memweave.config import MemoryConfig, EmbeddingConfig

config = MemoryConfig(
    embedding=EmbeddingConfig(
        model="ollama/nomic-embed-text",
        api_base="http://localhost:11434",
    )
)

Keyword-only mode (fully offline, no embeddings):

from memweave.config import MemoryConfig, QueryConfig

config = MemoryConfig(
    query=QueryConfig(strategy="keyword")
)

Issues and pull requests are welcome. Please open an issue before starting large changes.

🦞 OpenClaw — the memory architecture that inspired memweave.

MIT