remembox 0.2.1

Production-grade plug-and-play memory for AI agents. Zero config, single-file storage, framework-agnostic.

remembox

Production-grade plug-and-play memory for AI agents.

Give any LLM, agent, or rule-based system persistent episodic + semantic memory. Zero config, single-file storage, framework-agnostic.

from remembox import Remembox

memory = Remembox("my_agent.db")
memory.record("User said they love hiking in the Himalayas", importance=0.8)
memory.learn("user", "prefers", "black coffee", confidence=0.9)

# Retrieve relevant memories
results = memory.recall("what are the user's hobbies?", k=3)

# Get prompt-ready context string
context = memory.context("what does the user like?")
# → "## User Profile\n- user prefers black coffee (90%)\n\n## Relevant Memories\n- ..."

Features

Feature	Description
Episodic memory	Timestamped events with importance + emotion scoring
Semantic memory	Fact triples with reinforcement & contradiction
Procedural memory	Trigger → action routines, surfaced in context
Temporal facts	valid_from / valid_until windows + recurrence patterns
Reflection	Synthesizes higher-order patterns across episodes
Smart retrieval	Recency × relevance × importance (keyword or embeddings)
Context builder	Token-budgeted prompt string: profile + procedures + memories + patterns
Thread summarization	pi-style compaction of long conversation threads
Consolidation	Compress episodes → stable facts
Forgetting	Importance-weighted decay (trivial memories die, critical ones survive)
Editing & correction	In-place episode edits + timestamped annotation audit trail
Multi-user isolation	owner_id scoping — one DB, many users, no cross-contamination
Embeddings (optional)	SQLite-backed semantic retrieval via sentence-transformers
Zero dependencies	Core runs on Python stdlib only (SQLite)
Single-file storage	One .db file = entire memory. Copy it, back it up, version it.

Install

# From source (development)
git clone <repo-url> && cd remembox
uv sync --extra dev

# Coming soon: pip install remembox

Quick Start

Want to run these examples in a browser? Open notebooks/quickstart.ipynb — a 5-minute executable version of everything below. For the full simple→advanced tour, see notebooks/walkthrough.ipynb.

1. Record Events

from remembox import Remembox

memory = Remembox("jarvis.db")

# Record what happens
memory.record("User got promoted to Director of Engineering!", importance=1.0, emotion="ecstatic")
memory.record("User ordered black coffee", importance=0.3)
memory.record("User's dog Rocky had a vet visit. All clear.", importance=0.5, emotion="relieved")

2. Learn Facts

# Semantic facts with automatic conflict resolution
memory.learn("user", "name", "Pranav", confidence=0.95)
memory.learn("user", "prefers", "black coffee", confidence=0.9)
memory.learn("user", "lives_in", "Delhi", confidence=0.8)

# Repeat a fact → confidence increases (reinforcement)
memory.learn("user", "prefers", "black coffee")  # → confidence: 0.9 → 0.915

# New info → old fact deactivated (contradiction)
memory.learn("user", "lives_in", "Mumbai")  # Delhi deactivated, Mumbai active

3. Retrieve Memories

# Keyword + recency + importance scoring
results = memory.recall("coffee", k=3)
for r in results:
    print(f"Score: {r.score:.3f} | {r.episode.content}")
    # Score breakdown: R=recency, V=relevance, I=importance
    print(f"  R={r.recency:.2f} V={r.relevance:.2f} I={r.importance:.2f}")

4. Build LLM Context

# Generate a prompt-ready context string
context = memory.context("What should I get for dinner?", max_tokens=2000)
print(context)
# Output:
# ## User Profile
# - user name Pranav (95%)
# - user prefers black coffee (92%)
# - user lives_in Mumbai (80%)
#
# ## Relevant Memories
# - (2h ago) User ordered black coffee
# - (1d ago [ecstatic]) User got promoted to Director of Engineering!
# (when procedures or reflections exist, `## Active Procedures` and `## Patterns`
#  sections are added too)

# Plug into any LLM:
messages = [
    {"role": "system", "content": f"You are a helpful assistant.\n\n{context}"},
    {"role": "user", "content": "What should I get for dinner?"},
]
# → Send to OpenAI, Anthropic, Ollama, etc.

5. Maintenance

# Compress old episodes into semantic facts
memory.consolidate()        # one batch; use consolidate_all() to drain the backlog

# Synthesize higher-order patterns across episodes
memory.reflect()

# Prune stale memories (importance-weighted; critical ones survive)
result = memory.forget()
print(f"Deleted: {result['deleted']}, Archived: {result['archived']}")

# …or run all of the above in one call:
memory.maintain()

# Health check
print(memory.stats())

Configuration

from remembox import Remembox, MemoryConfig

# Custom config
config = MemoryConfig(
    decay_rate=0.05,          # How fast memories fade (higher = faster)
    w_recency=0.3,            # Retrieval weight: recency
    w_relevance=0.4,          # Retrieval weight: keyword match
    w_importance=0.3,         # Retrieval weight: stored importance
    max_context_tokens=2000,  # Token budget for context()
)
memory = Remembox("agent.db", config=config)

# Or use presets:
fast_memory = Remembox("chatbot.db", config=MemoryConfig.fast())    # Aggressive forgetting
deep_memory = Remembox("assistant.db", config=MemoryConfig.deep())  # Long retention

Integration Examples

With OpenAI

from openai import OpenAI
from remembox import Remembox

client = OpenAI()
memory = Remembox("openai_agent.db")

def chat(user_message: str) -> str:
    # Store the user message
    memory.record(f"User: {user_message}", importance=0.5)

    # Build context-enhanced prompt
    context = memory.context(user_message)
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": f"You are a helpful assistant.\n\n{context}"},
            {"role": "user", "content": user_message},
        ],
    )
    reply = response.choices[0].message.content

    # Store the response
    memory.record(f"Assistant: {reply}", importance=0.2, source="response")
    return reply

With Anthropic

import anthropic
from remembox import Remembox

client = anthropic.Anthropic()
memory = Remembox("claude_agent.db")

def chat(user_message: str) -> str:
    memory.record(f"User: {user_message}")
    context = memory.context(user_message)

    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        system=f"You are a helpful assistant.\n\n{context}",
        messages=[{"role": "user", "content": user_message}],
        max_tokens=1024,
    )
    reply = response.content[0].text
    memory.record(f"Assistant: {reply}", importance=0.2, source="response")
    return reply

With Any Agent Framework

from remembox import Remembox

memory = Remembox("agent.db")

# In your agent's observe/act loop:
def agent_step(observation: str) -> str:
    # 1. Store observation
    memory.record(observation, importance=score_importance(observation))

    # 2. Retrieve relevant context
    context = memory.context(observation)

    # 3. Generate action (your logic here)
    action = your_llm_or_rules(observation, context)

    # 4. Periodic maintenance
    if should_consolidate():
        memory.consolidate()
        memory.forget()

    return action

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      Remembox                            │
│                                                              │
│  record()  recall()  learn()  context()  consolidate()      │
│     │         │        │         │            │              │
│     ▼         ▼        ▼         ▼            ▼              │
│  ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐   │
│  │Episodic│ │Retriev.│ │Semantic│ │Context │ │Consol. │   │
│  │ Store  │ │ Engine │ │ Store  │ │Builder │ │Pipeline│   │
│  └───┬────┘ └────────┘ └───┬────┘ └────────┘ └────────┘   │
│      │                      │                               │
│      └──────────┬───────────┘                               │
│                 ▼                                            │
│           ┌──────────┐                                      │
│           │  SQLite   │  ← Single .db file                  │
│           │  (WAL)    │                                      │
│           └──────────┘                                      │
└─────────────────────────────────────────────────────────────┘

API Reference

Grouped by memory type. All methods are on Remembox.

Category	Method	Returns
Episodic	record(content, importance, emotion, source, context, thread_id, parent_id, depth)	Episode
Episodic	recall(query, k, min_score)	list[RetrievalResult]
Episodic	recent(n)	list[Episode]
Episodic	search(keyword, limit)	list[Episode]
Threads	thread(thread_id, limit)	list[Episode]
Threads	thread_children(episode_id, limit)	list[Episode]
Threads	threads(limit)	list[str]
Threads	summarize_thread(thread_id)	ThreadSummaryResult
Semantic	learn(subject, predicate, obj, confidence, valid_from, valid_until, recurrence)	(Fact, action)
Semantic	about(subject, at_time)	list[Fact]
Semantic	find_fact(subject, predicate, at_time)	list[Fact]
Procedural	learn_procedure(trigger, action, confidence)	Procedure
Procedural	match_procedures(text)	list[Procedure]
Procedural	procedures() / delete_procedure(id)	list[Procedure] / bool
Context	context(query, max_tokens, min_score, profile_subject)	str
Maintenance	consolidate() / consolidate_all()	dict
Maintenance	reflect(episodes) / reflections(subject)	dict / list
Maintenance	forget()	dict
Maintenance	maintain() — runs consolidate → reflect → summarize → forget	dict
Editing	update_episode(id, ...) / annotate_episode(id, ...)	Episode
Editing	edit_fact(id, ...) / correct_fact(id, ...)	Fact / (Fact, action)
Lifecycle	stats() / close()	dict / None

learn()'s action is 'new' / 'reinforced' / 'contradicted'. Retrieval scores combine as w_recency·R + w_relevance·V + w_importance·I (see notebooks/walkthrough.ipynb §5).

Design Decisions

Decision	Why
SQLite, not Postgres	Zero config, ~0.05ms reads, single-file backup. Handles 18M rows (10yr heavy use).
Stdlib only	No supply-chain risk. Embedding search is an optional extra.
Sync API	Lowest latency path. SQLite calls are <1ms — async overhead isn't worth it.
Store protocols	EpisodicStoreProtocol / SemanticStoreProtocol — swap in Postgres/Redis by implementing the protocol.
Tiered forgetting	Critical memories (importance ≥ 0.9) never die. Trivial ones fade in days.

Development

# Install dev deps
uv sync --extra dev

# Run tests (262 passed, 19 skipped, ~1.5s)
uv run pytest

# Run with verbose output
uv run pytest -v

# Run scale tests only
uv run pytest tests/test_episodic.py::TestEpisodicScale -v

Documentation

Doc	What it's for
README.md (this file)	Front door: features, install, quick start, API, design
notebooks/	Runnable examples — quickstart.ipynb (5-min) and walkthrough.ipynb (simple→advanced)
CHANGELOG.md	History of changes, release by release
BUGS.md	Known-issues tracker — audit with severity, status, and fix notes
lessons/	8 incremental teaching scripts that build the library from scratch
demos/	demo_openrouter.py — live end-to-end run with a real LLM
integrations/pi/	Long-term memory for pi: HTTP sidecar + TypeScript extension

License

MIT