widemem.ai

.__    .___                                        .__
__  _  _|__| __| _/____   _____   ____   _____      _____  |__|
\ \/ \/ /  |/ __ |/ __ \ /     \_/ __ \ /     \     \__  \ |  |
 \     /|  / /_/ \  ___/|  Y Y  \  ___/|  Y Y  \     / __ \|  |
  \/\_/ |__\____ |\___  >__|_|  /\___  >__|_|  / /\ (____  /__|
                \/    \/      \/     \/      \/  \/      \/

>   *Goldfish memory? ¬_¬ Fixed.*

Background reading:

• Whitepaper: How LLMs Handle Memory. Technical paper on memory architectures, security risks, and in-weights personalisation.
• Why Context Windows Aren't Memory. The problem widemem solves.
• Your AI Memory Can't Tell a River Bank from a Savings Account. How YMYL classification actually works.
• Your AI Should Know When It Doesn't Know. Uncertainty-aware retrieval.
• Corrections log. Published claims that turned out wrong, and their fixes.

Because your AI deserves better than amnesia. ¬_¬

An open-source AI memory layer that actually remembers what matters. Local-first, batteries-included, and opinionated about not forgetting your user's blood type.

Look, AI memory has come a long way. Context windows are bigger, RAG pipelines are everywhere, and most frameworks have some form of "remember this for later." It's not terrible anymore. But it's not great either. Most memory systems treat every fact the same: your user's blood type sits next to what they had for lunch, decaying at the same rate, with the same priority. Contradictions pile up silently. There's no sense of "this matters more than that." And when you need to remember something from three months ago that actually matters? Good luck.

widemem is for when "good enough" isn't good enough.

widemem gives your AI a real memory: one that scores what matters, forgets what doesn't, and absolutely refuses to lose track of someone's prescription medication just because 72 hours passed and the decay function got bored. Think of it as long-term memory for LLMs, except it actually works and doesn't require a PhD to set up.

• Memories that know their place. Importance scoring (1-10) plus time decay means "has a peanut allergy" always outranks "had pizza on Tuesday". As it should. Not all memories are created equal, and your retrieval system should know the difference between a life-threatening allergy and a lunch preference.
• One brain, three layers. Facts roll up into summaries, summaries into themes. Ask "where does Alice live" and get the fact. Ask "tell me about Alice" and get the big picture. Your AI can zoom in and zoom out without breaking a sweat or making a second API call.
• YMYL or GTFO. Health, legal, and financial facts get VIP treatment: higher importance floors, immunity from decay, and forced contradiction detection. Two-stage classification (regex for obvious matches, LLM for implied content) catches "my chest hurts" as health while ignoring "the bank of the river." Read more ↗
• Conflict resolution that isn't stupid. Add "I live in Boston" after "I live in San Francisco" and the system doesn't just blindly append both. It detects the contradiction, resolves it in a single LLM call, and updates the memory. Like a reasonable adult would.
• Graceful memory-miss handling. Every retrieval returns a confidence level (HIGH / MODERATE / LOW / NONE) so your agent knows when memory has nothing relevant and can abstain instead of guessing. Three modes: `strict` (refuse on low confidence), `helpful` (hedge with related context), `creative` (offer to guess, with a warning). For high-stakes contexts where a wrong answer is worse than no answer.
• Local by default, cloud if you want. SQLite plus FAISS out of the box. No accounts, no API keys for storage, no "please sign up for our enterprise plan to store more than 100 memories". Plug in Qdrant or any cloud provider when you're ready. Or don't. We won't guilt-trip you.

Architecture

TL;DR

Seven features, one library. Here's what widemem does that most memory systems don't:

600+ tests. Zero external services required. SQLite plus FAISS by default. Plug in OpenAI, Anthropic, Ollama, Qdrant, or sentence-transformers as needed.

Table of Contents

• Install
• Quick Start
• Configuration
• Scoring & Decay
• Providers
• YMYL (Your Money or Your Life)
• Hierarchical Memory
• Active Retrieval
• Temporal Search
• Uncertainty & Confidence
• Retrieval Modes
• History & Audit Trail
• Batch Conflict Resolution
• Prompt-Injection Sanitizer
• API Reference
• Claude Code Skill
• MCP Server
• Development
• Benchmarks
• Disclaimer & intended use
• Contact
• License

Install

pip install widemem-ai[faiss]

The `[faiss]` extra installs the default local vector store. Plain `pip install widemem-ai` installs the core only; you'll need at least one vector backend (`[faiss]` or `[qdrant]`) before `WideMemory()` will work. Python 3.10+ required.

Optional providers

pip install widemem-ai[anthropic]             # Claude LLM provider
pip install widemem-ai[ollama]                # Local LLM via Ollama
pip install widemem-ai[sentence-transformers] # Local embeddings (no API key needed)
pip install widemem-ai[qdrant]                # Qdrant vector store
pip install widemem-ai[mcp]                   # Model Context Protocol server
pip install widemem-ai[all]                   # Everything. You want it all? You got it.

Quick Start

Five lines to a working memory system. Six if you count the import.

from widemem import WideMemory, MemoryConfig

memory = WideMemory()

# Add memories
result = memory.add("I live in San Francisco and work as a software engineer", user_id="alice")

# Search
results = memory.search("where does alice live", user_id="alice")
for r in results:
    print(f"{r.memory.content} (score: {r.final_score:.2f})")

# Update happens automatically. Add contradicting info and the resolver handles it.
memory.add("I just moved to Boston", user_id="alice")

# Delete
memory.delete(results[0].memory.id)

# History audit trail
history = memory.get_history(results[0].memory.id)

That's it. No 47-step setup guide. No YAML files. No existential dread. Your AI just went from goldfish to elephant in six lines.

WideMemory also works as a context manager if you're the responsible type:

with WideMemory() as memory:
    memory.add("I live in San Francisco", user_id="alice")
    results = memory.search("where does alice live", user_id="alice")
# Connection closed automatically. You're welcome.

Configuration

Most defaults are sane, so a minimal config is usually enough:

from widemem import WideMemory, MemoryConfig
from widemem.core.types import LLMConfig, ScoringConfig, YMYLConfig

config = MemoryConfig(
    llm=LLMConfig(provider="openai", model="gpt-4o-mini"),
    scoring=ScoringConfig(decay_rate=0.01),
    ymyl=YMYLConfig(enabled=True),
    history_db_path="~/.widemem/history.db",
)
memory = WideMemory(config)

Full reference for every field, default, and tradeoff: **docs/configuration.md**.

Scoring & Decay

The Formula

Every search result gets a combined score. It's not rocket science, but it's close enough:

final_score = (similarity_weight * similarity) + (importance_weight * importance) + (recency_weight * recency)
final_score *= topic_boost   # if topic weights are set

• `similarity`: cosine similarity from vector search (0-1)
• `importance`: normalized from the 1-10 rating assigned at extraction (0-1)
• `recency`: time decay score (0-1), computed by the decay function
• `topic_boost`: multiplier from topic weights (default 1.0)

Decay Functions

Control how memories fade over time. Like real memories, but configurable. Unlike a goldfish, you can turn decay off entirely.

# Fast decay: what happened last week? who cares
ScoringConfig(decay_function=DecayFunction.EXPONENTIAL, decay_rate=0.05)

# Slow decay: memories stay relevant longer
ScoringConfig(decay_function=DecayFunction.EXPONENTIAL, decay_rate=0.005)

# No decay: all memories equally fresh forever
ScoringConfig(decay_function=DecayFunction.NONE)

Providers

For Ollama, pair with sentence-transformers if you want fully local: `EmbeddingConfig(provider="sentence-transformers", model="all-MiniLM-L6-v2", dimensions=384)`. Set `QDRANT_URL` env var for remote Qdrant.

YMYL (Your Money or Your Life)

Some facts are more equal than others. YMYL prioritization ensures that critical facts about health, finances, legal matters, and safety are never lost, never deprioritized, and never quietly forgotten because the decay function decided Tuesday was a good day to forget someone's insulin dosage.

> For the full deep dive on how YMYL works, edge cases, and limitations, see **YMYL.md**.

config = MemoryConfig(
    ymyl=YMYLConfig(
        enabled=True,
        categories=["health", "medical", "financial", "legal", "safety", "insurance", "tax", "pharmaceutical"],
        min_importance=8.0,          # Floor importance for strong YMYL facts
        decay_immune=True,           # Strong YMYL facts don't decay over time
        force_active_retrieval=True, # Force contradiction detection for strong YMYL facts
    ),
)

Two-Stage Semantic Classification

Not every mention of "bank" means someone's talking about their finances. And "my chest has been hurting for three days" is a health concern even though it contains no medical keyword. widemem uses a two-stage pipeline to handle both cases:

Strong regex matches get immediate YMYL protection. For everything else, the LLM decides based on context. This catches implied YMYL content ("I stopped taking my pills" -> medical) and rejects false positives ("The Doctor is a great TV show" -> not medical).

> For the full breakdown with accuracy data and examples, see **Your AI Memory Can't Tell a River Bank from a Savings Account**.

YMYL Categories

8 categories, each with strong (unambiguous) and weak (context-dependent) patterns:

You can enable a subset if you only care about some categories:

YMYLConfig(enabled=True, categories=["health", "medical", "financial"])

Topic Weights (related)

Boost or suppress specific topics during retrieval as a multiplier on `final_score`:

config = MemoryConfig(
    topics=TopicConfig(
        weights={"python": 2.0, "cooking": 0.5},
        custom_topics=["python", "machine learning"],  # Extraction hints
    ),
)

Matching is case-insensitive substring. Values above 1.0 boost, below 1.0 suppress. `custom_topics` are passed to the LLM during extraction as a hint.

Hierarchical Memory

Three-tier memory system. Facts are great, but sometimes you need the big picture.

config = MemoryConfig(enable_hierarchy=True)
memory = WideMemory(config)

# Add many facts
for msg in conversation_history:
    memory.add(msg, user_id="alice")

# Trigger summarization (groups related facts, creates summaries and themes)
memory.summarize(user_id="alice")

# Broad queries return themes, specific queries return facts
results = memory.search("tell me about alice")        # Returns themes
results = memory.search("where does alice live")      # Returns facts

# Filter by tier
from widemem.core.types import MemoryTier
results = memory.search("alice", tier=MemoryTier.SUMMARY)

Tiers

Query routing uses keyword heuristics (no extra LLM call) with a fallback chain. If the preferred tier has no results, it falls back to the next tier. No results left behind.

Active Retrieval

Your AI shouldn't silently overwrite "lives in San Francisco" with "lives in Boston" without at least raising an eyebrow. Active retrieval detects contradictions and ambiguities, then asks clarifying questions via callbacks. Read more ↗

config = MemoryConfig(
    enable_active_retrieval=True,
    active_retrieval_threshold=0.6,  # Similarity threshold for conflict detection
)
memory = WideMemory(config)

def handle_clarification(clarifications):
    for c in clarifications:
        print(f"Conflict: {c.question}")
        print(f"  Old: {c.existing_memory}")
        print(f"  New: {c.new_fact}")
    # Return None to abort the add, or a list of answers to proceed
    return ["User moved to Boston"]

result = memory.add(
    "I just moved to Boston",
    user_id="alice",
    on_clarification=handle_clarification,
)

if result.has_clarifications:
    print(f"Resolved {len(result.clarifications)} conflicts")

Callback behavior

• `on_clarification` receives a list of `Clarification` objects
• Return `None` to abort the add entirely (the nuclear option)
• Return a list of strings (answers) to proceed with the add
• If no callback is provided, the add proceeds and clarifications are returned in `AddResult.clarifications` for you to deal with later. Or never. We won't judge.

Temporal Search

Filter and rank memories by time. Because sometimes you only care about what happened recently.

from datetime import datetime, timedelta

now = datetime.utcnow()

# Only memories from the last week
results = memory.search(
    "what happened recently",
    user_id="alice",
    time_after=now - timedelta(days=7),
)

# Only memories before January 2026
results = memory.search(
    "old preferences",
    user_id="alice",
    time_before=datetime(2026, 1, 1),
)

# Combined range
results = memory.search(
    "december events",
    user_id="alice",
    time_after=datetime(2025, 12, 1),
    time_before=datetime(2025, 12, 31),
)

Uncertainty & Confidence

Every retrieval returns a `RetrievalConfidence` level (`HIGH`, `MODERATE`, `LOW`, `NONE`) based on how relevant the top results are. Your agent can use this to abstain on low-confidence queries instead of guessing from irrelevant memories. Three response modes (`strict`, `helpful`, `creative`) let you tune the abstention behavior to the use case. Read more ↗

Every search returns a confidence level:

response = mem.search("What's Alice's favorite movie?", user_id="alice")

response.confidence     # RetrievalConfidence.NONE: nothing relevant found
response.has_relevant   # False

# But it still works like a list (backward compatible):
for r in response:
    print(r.memory.content)

Three uncertainty modes

# Strict: refuses to answer if unsure
mem = WideMemory(config=MemoryConfig(uncertainty_mode="strict"))

# Helpful (default): "I don't have that, but here's what I do know..."
mem = WideMemory(config=MemoryConfig(uncertainty_mode="helpful"))

# Creative: "I can guess if you want, fair warning, it might be wrong"
mem = WideMemory(config=MemoryConfig(uncertainty_mode="creative"))

Pin important memories

When a user explicitly tells you something important, pin it so it sticks:

# Normal add: importance decided by LLM (might be 3-6)
mem.add("I had pasta for lunch", user_id="alice")

# Pin: stored with importance 9, resistant to decay
mem.pin("My blood type is O negative", user_id="alice")

Frustration recovery

When users say "I told you this!", widemem detects the frustration, extracts the fact, and offers to pin it:

from widemem.retrieval.uncertainty import build_frustration_response

response = build_frustration_response(
    "I told you my blood type is O negative!",
    confidence=RetrievalConfidence.NONE,
    mode=UncertaintyMode.HELPFUL,
)
# response = {
#     "action": "recover_and_pin",
#     "message": "Sorry about that. I'm saving this now with high importance.",
#     "pin_fact": "my blood type is O negative",
#     "pin_importance": 9.0,
# }

Retrieval Modes

Not every query needs the same depth. A casual chatbot doesn't need 50 retrieved memories. A medical assistant does. widemem lets you choose:

from widemem import WideMemory, MemoryConfig, RetrievalMode

# Set at config level (default for all queries)
mem = WideMemory(config=MemoryConfig(retrieval_mode="balanced"))

# Override per query when needed
results = mem.search("critical question", mode=RetrievalMode.DEEP)

Each mode also adjusts the internal candidate pool size and similarity boost strength. `balanced` is the sweet spot for most use cases. Enough context for good answers without burning tokens.

History & Audit Trail

Every add, update, and delete is logged to SQLite. Full audit trail. Because "who changed this and when" is a question you'll eventually ask.

history = memory.get_history(memory_id)
for entry in history:
    print(f"{entry.timestamp}: {entry.action.value}")
    if entry.old_content:
        print(f"  From: {entry.old_content}")
    if entry.new_content:
        print(f"  To: {entry.new_content}")

Batch Conflict Resolution

When new facts are added, widemem finds related existing memories and sends everything to the LLM in a single call. The LLM decides for each fact whether to ADD (new), UPDATE (modify existing), DELETE (contradicted), or NONE (duplicate).

This is the main architectural improvement over per-fact approaches. One call instead of N. The LLM sees the full context and can make better decisions. Your API bill sees fewer line items.

Prompt-Injection Sanitizer

Memory content gets fed back into LLM prompts at extraction, conflict resolution, summarization, and answer time. Hostile content stored once can poison every later call. widemem strips well-known prompt-injection patterns before content reaches the LLM:

• Direct instruction overrides (`ignore previous instructions`, `disregard the rules`, `forget what I said`)
• System-prompt tags (``, ``, `[system]`)
• Role markers at line start (`system:`, `assistant:`)
• Common jailbreak vocabulary (`DAN mode`, `developer mode`)
• Memory-targeted destructive actions (`delete all memories`)

Conservative by design: only the most well-established attack patterns are matched, so legitimate clinical or operational content like "ignore all previous medications" or "the patient often forgets everything by morning" passes through untouched.

from widemem.security import detect_injection, sanitize

cats = detect_injection("Please ignore all previous instructions.")
# ["instruction-override"]

sanitized, found = sanitize("do harmful stuff")
# sanitized = "[REDACTED]do harmful stuff[REDACTED]"
# found = ["system-tag", "system-tag"]

The sanitizer runs automatically inside `LLMExtractor.extract()`. This is a baseline defense, not a complete solution: defense-in-depth still requires output validation, structured prompts that distinguish data from instruction, and provider-side guardrails.

Self-Supervised Extraction

widemem can collect extraction training pairs (`collect_extractions=True` in `MemoryConfig`) and let you distill a small local model from them, falling back to the LLM when the small model's confidence is low. Code in `widemem/extraction/collector.py`. Training scripts under `scripts/`.

Collection is off by default and opt-in, because it persists raw, pre-sanitization input text (a PII risk). `ExtractionCollector` stays disabled unless you pass `enabled=True` or set `WIDEMEM_COLLECT_EXTRACTIONS=1`; while disabled it opens no database and every operation is a no-op.

API Reference

Full method signatures, parameters, and return types: **docs/api.md**.

The most-used surface area:

Claude Code Skill

Try widemem directly in Claude Code with the official memory skill.

Install

pip install widemem-ai[mcp,sentence-transformers]

Available commands

Skill repo

Full setup instructions and source: widemem-skill.

MCP Server

widemem ships an MCP server for Claude Desktop, Cursor, or any MCP-compatible client.

pip install widemem-ai[mcp]
python -m widemem.mcp_server

Tools exposed: `widemem_add`, `widemem_search`, `widemem_delete`, `widemem_count`, `widemem_health`. Configure providers via `WIDEMEM_LLM_PROVIDER`, `WIDEMEM_EMBEDDING_PROVIDER`, etc.

Full setup, env vars, and Claude Desktop config: **docs/mcp.md**.

Development

git clone https://github.com/remete618/widemem-ai
cd widemem-ai
pip install -e ".[dev,faiss]"
pytest

600+ tests. They all pass. We checked.

Benchmarks

Measured on the full 1,540-question LoCoMo benchmark, v1.5.0:

Mid-pack accuracy at a fraction of the token cost: reference systems spend 1,700 to 26,000 tokens per query. Per-category labels published before 2026-07-06 had single-hop and multi-hop transposed; the multi-hop leadership claim is retracted and the correction is logged in docs/HISTORY.md. Full methodology, per-category breakdowns, and reference-system comparisons: widemem.ai/benchmarks.

To re-run it: the harness (`benchmark/run_ws1.py`, `val.py`, `honest_core.py`) and the question split (`benchmark/locomo_split.json`) are in this repo. The LoCoMo dataset itself is not vendored here, so fetch it from snap-research/locomo into `benchmark/locomo-data/` first. Published result files are not committed.

Roadmap

Tracked publicly as GitHub issues. Vote with reactions to prioritize. Issues tagged `good first issue` are ideal entry points for new contributors. Each carries a scope, a quality bar, and a 48-hour review SLA in the body.

Audit-grade core

• #21 Source-message provenance — link every fact in the history log back to the inbound message that produced it

Framework integrations

• #22 LangChain `BaseChatMessageHistory` adapter — drop-in conversation-history backend for LangChain chains and agents
• #23 LangChain `BaseRetriever` adapter — RAG-style retrieval from widemem in any LangChain chain
• #24 LangGraph `BaseStore` adapter — memory backend for stateful LangGraph agents

In flight

• #6 Streaming memory search — async iterator over results as they rank (claimed by @harishkotra)

What we are explicitly not building: 20-provider integration matrix, additional vector store backends beyond FAISS and Qdrant, hosted multi-tenant service, web UI for memory management, GraphQL API, command-line interface. The 80/20 is the audit-grade core for regulated deployments. Everything else is application code.

Disclaimer & intended use

widemem is developer infrastructure, provided under the Apache License 2.0, as is and without warranty of any kind. It is not medical, legal, tax, or financial advice, not a medical device, and not a substitute for a qualified professional. Its YMYL handling (regex plus LLM classification) is a best-effort safety net, not a guarantee. Keep a human in the loop and verify outputs before relying on them in any high-stakes decision.

You are responsible for your own deployment, the data you store, and meeting the regulatory obligations that apply to you. When you self-host, your data stays in your environment and we receive nothing.

Export note: the software may be subject to export-control and sanctions laws (including the US EAR and OFAC lists). Do not download, use, or re-export it in violation of those laws.

The Apache 2.0 license in LICENSE governs your use of the code. Terms for the hosted service and website are at widemem.ai/terms. LLM provider terms apply to provider API calls.

Contact

• Email: hello@widemem.ai
• Project: widemem.ai
• Repository: github.com/remete618/widemem-ai

Bug reports, feature requests, and unsolicited opinions are all welcome at the GitHub issues page.

License

Apache 2.0. See LICENSE for the full text that nobody reads.