recollectx 0.2.2

Belief-centric memory layer for agentic AI systems

re-collect

A belief-centric memory layer for agentic AI systems.

re-collect structures memory as beliefs — not logs. Every piece of information is stored as a typed claim with a confidence score, supporting evidence, and explicit relationships to other beliefs. This makes memory inspectable, revisable, and suitable for agents that need to reason about what they know and why.

---

Features

• Structured beliefs — two claim types: SemanticClaim (facts as subject-predicate-object triples) and EpisodicClaim (events with a summary)
• Confidence scoring — every claim has a confidence value in [0.0, 1.0]; contradictions decay it, supports boost it
• Belief graph — tracks supports, contradicts, derives, and similar relationships between beliefs
• Write policies — composable filters that accept or reject claims before storage
• LLM-powered updates — intelligent ADD / UPDATE / DELETE decisions using any LLM provider
• Pluggable storage — SQLite (SQLAlchemy) + optional vector backends (FAISS, Qdrant, Pinecone)
• LangGraph agent — answer questions by retrieving from memory using tool-based reasoning
• Automatic extraction — extract structured claims from raw text via LLM
• Deduplication — detect and merge duplicate claims using embedding similarity + LLM merge decisions
• Deep explanations — trace why a belief exists with recursive justification chains

---

Installation

Core (SQLite storage only)

pip install re-collect

With optional extras

# OpenAI LLM + embeddings
pip install "re-collect[openai]"

# Anthropic Claude
pip install "re-collect[anthropic]"

# Local Ollama models
pip install "re-collect[ollama]"

# Qdrant vector backend
pip install "re-collect[qdrant]"

# Pinecone vector backend
pip install "re-collect[pinecone]"

# Local FAISS + sentence-transformers (no API needed)
pip install "re-collect[local]"

# All LLM providers
pip install "re-collect[llm]"

# All vector backends
pip install "re-collect[vector]"

# Everything
pip install "re-collect[all]"

---

Quick Start

from recollectx import Memory, SemanticClaim, EpisodicClaim
from recollectx.db import SessionLocal, create_tables
from recollectx.storage import MemoryStore
from recollectx.storage.vector import FAISSBackend

# 1. Set up the database
create_tables()
db = SessionLocal()

# 2. Set up a vector backend (requires re-collect[local])
vectors = FAISSBackend(embed_fn=my_embed_fn, dimension=384)
store = MemoryStore(db, vectors)

# 3. Create a Memory instance
memory = Memory(storage=store)

# 4. Store beliefs
fact = SemanticClaim(
    subject="sky",
    predicate="has_color",
    object="blue",
    confidence=0.9,
    evidence=("direct observation",),
)
memory.store(fact)

event = EpisodicClaim(
    summary="User mentioned they love pizza on Fridays",
    confidence=0.85,
    importance=0.7,
)
memory.store(event)

# 5. Retrieve beliefs
facts = memory.retrieve(type="semantic")
episodes = memory.retrieve(type="episodic")

# 6. Explain a belief
explanation = memory.explain(fact.id)
print(explanation["supported_by"])    # belief IDs that support this fact
print(explanation["contradicted_by"]) # belief IDs that contradict it

---

Core Concepts

Claims

Claims are immutable dataclasses representing beliefs:

from recollectx import SemanticClaim, EpisodicClaim

# A fact: subject → predicate → object
fact = SemanticClaim(
    subject="user",
    predicate="prefers",
    object="dark mode",
    confidence=0.95,
    importance=0.8,        # 0.0 = trivial, 1.0 = critical
    evidence=("user said so in chat",),
)

# An event
event = EpisodicClaim(
    summary="User completed onboarding on March 28",
    confidence=0.99,
    importance=0.6,
)

Common fields on all claims:

Field	Type	Description
id	str	Auto-generated UUID
confidence	float	How certain is this belief? [0.0, 1.0]
importance	float	How useful for future conversations? [0.0, 1.0]
evidence	tuple[str, ...]	Supporting evidence strings
support_count	int	Times this belief has been reinforced

---

Memory

Memory is the central interface. It wires together storage, write policies, the belief graph, optional LLM updater, and confidence propagation:

from recollectx import Memory
from recollectx.policies import MinConfidence, MinEvidence

memory = Memory(
    storage=store,
    write_policy=MinConfidence(0.5) & MinEvidence(1),  # composable policies
)

Key methods:

Method	Description
store(claim)	Store a claim (applies policy first)
retrieve(**kwargs)	Query claims by type, confidence, etc.
explain(belief_id)	Get direct supports/contradictions for a belief
explain_deep(belief_id, max_depth)	Recursive justification chain
explain_confidence_history(belief_id)	Full confidence change history
add_support(src_id, dst_id)	Manually add a support relationship

---

Belief Graph

The belief graph tracks relationships between claims:

# Relationships are created automatically by the LLM updater,
# or you can add them manually:
memory.add_support(src_id=claim_a.id, dst_id=claim_b.id)

# Deep explanation traverses the graph
result = memory.explain_deep(claim_b.id, max_depth=3)
print(result.root.belief)        # the target belief
print(result.total_nodes)        # how many nodes in the explanation
print(result.cycle_detected)     # True if circular reasoning was found

---

Write Policies

Policies filter claims before they reach storage. Combine with &:

from recollectx.policies import MinConfidence, MinEvidence

# Only store claims with confidence >= 0.6 and at least 1 piece of evidence
policy = MinConfidence(0.6) & MinEvidence(1)

memory = Memory(storage=store, write_policy=policy)

Custom policies implement a simple callable protocol:

from recollectx.policies.base import Decision

class MyPolicy:
    def __call__(self, claim, memory) -> Decision:
        if claim.importance < 0.3:
            return Decision.REJECT
        return Decision.ACCEPT

---

Confidence Propagation

When beliefs are related, confidence flows between them automatically:

from recollectx import Memory, PropagationConfig

config = PropagationConfig(
    support_boost=0.10,        # supports increase confidence by this amount
    contradiction_decay=0.15,  # contradictions reduce confidence by this amount
)

memory = Memory(storage=store, propagation_config=config)
# Now adding support/contradiction edges will adjust confidence scores

---

LLM Integration

LLM Providers

from recollectx.llm.providers import OpenAIProvider, AnthropicProvider, OllamaProvider

# OpenAI (requires re-collect[openai])
llm = OpenAIProvider(api_key="sk-...", model="gpt-4o-mini")

# Anthropic (requires re-collect[anthropic])
llm = AnthropicProvider(api_key="sk-ant-...", model="claude-3-5-haiku-latest")

# Ollama — local, free (requires re-collect[ollama])
llm = OllamaProvider(model="llama3", base_url="http://localhost:11434")

LLM-Powered Memory Updates

The MemoryUpdater uses an LLM to make intelligent write decisions:

from recollectx import Memory, MemoryUpdater

updater = MemoryUpdater(llm=llm, storage=store)
memory = Memory(storage=store, updater=updater)

# Now memory.store() will search for similar existing beliefs
# and decide: ADD, UPDATE, DELETE, or skip (NONE)
memory.store(new_claim)

Claim Extraction from Text

from recollectx.extractors import LLMExtractor

extractor = LLMExtractor(llm_provider=llm, min_confidence=0.5, max_claims_per_text=10)

claims = await extractor.extract("Alice loves hiking on weekends and prefers trail mix as a snack.")
for claim in claims:
    memory.store(claim)

---

Vector Backends

Vector backends enable semantic (similarity) search over beliefs.

FAISS (local, no server needed)

from recollectx.storage.vector import FAISSBackend

vectors = FAISSBackend(embed_fn=my_embed_fn, dimension=384)

Qdrant

from recollectx.storage.vector import QdrantBackend

vectors = QdrantBackend(
    url="http://localhost:6333",
    collection_name="beliefs",
    embedding_fn=my_embed_fn,
    distance="cosine",
)

Pinecone

from recollectx.storage.vector import PineconeBackend

vectors = PineconeBackend(
    api_key="your-api-key",
    index_name="beliefs",
    embedding_fn=my_embed_fn,
)

---

LangGraph Memory Agent

Answer questions by letting an agent retrieve from memory using tools:

from recollectx.agents import MemoryAgent
from langchain_ollama import ChatOllama

llm = ChatOllama(model="llama3")
agent = MemoryAgent(memory=memory, llm=llm)

response = agent.answer("What food does the user like?")
print(response.answer)       # "The user loves pizza"
print(response.tools_used)   # ["search_memories", "get_facts_about"]

Available retrieval tools: search_memories, get_recent_memories, get_facts_about, combine_facts.

---

Deduplication

Detect and merge duplicate or near-duplicate beliefs:

from recollectx.deduplication import ClaimDeduplicator

deduplicator = ClaimDeduplicator(
    storage=store,
    llm=llm,
    similarity_threshold=0.85,
)

await deduplicator.run()  # scans storage and merges duplicates

---

Architecture

re_collect/
├── claims.py          # EpisodicClaim, SemanticClaim dataclasses
├── memory.py          # Memory — main interface
├── updater.py         # LLM-powered write decisions
├── propagation.py     # Confidence propagation on relationships
├── state.py           # AgentState for dynamic state management
│
├── db/                # SQLAlchemy ORM (ClaimModel, BeliefEdgeModel, ConfidenceHistoryModel)
├── storage/           # MemoryStore (SQLite + vector), vector backends (FAISS, Qdrant, Pinecone)
├── graph/             # BeliefGraph, edges, deep explanation
├── policies/          # Composable write policies
├── llm/               # LLMProvider protocol + OpenAI/Anthropic/Ollama/OpenRouter/Local providers
├── extractors/        # LLM-powered claim extraction from text
├── agents/            # LangGraph-based Q&A agent + tools
└── deduplication/     # Embedding similarity + LLM merge decisions

---

Requirements

• Python >= 3.12
• SQLAlchemy >= 2.0
• LangChain ecosystem (langchain-core, langchain-openai, langchain-anthropic, langgraph)

Optional dependencies vary by feature — see Installation.

---

License

MIT