human-memory-sdk 0.1.1

Python SDK for the Human-Like Memory System — ACT-R based memory for AI agents

human-memory

Python SDK for the Human-Like Memory System — give your AI agents human-like memory based on the ACT-R cognitive architecture.

Unlike simple vector stores, this system mimics how the human brain actually works:

• Frequently used memories get stronger (activation from access history)
• Unused memories naturally fade (power-law decay)
• Emotional events persist longer (salience-modified decay)
• Context boosts related memories (spreading activation via knowledge graph)
• Repeated patterns become skills (consolidation extracts knowledge)

Installation

pip install human-memory

Requirements: Python 3.12+ | Only 2 dependencies: httpx, pydantic

Prerequisites

The SDK connects to a running human-memory server. Start it with:

git clone https://github.com/AbdulrahmanMasoud/human-memory.git
cd human-memory
docker compose up -d

The server runs at http://localhost:8000 by default.

---

Quick Start

from human_memory import Memory

# Connect to the memory server
memory = Memory("http://localhost:8000")

# Store a memory
result = memory.store("User prefers dark mode in all applications")
print(result.memory_id)  # "a1b2c3d4-..."
print(result.activation)  # 1.0 (initial activation)

# Search memories (ranked by ACT-R activation, not just similarity)
results = memory.search("what are the user's preferences?")
for r in results:
    print(f"  {r.content}")
    print(f"  activation: {r.activation:.2f} | similarity: {r.similarity:.2f}")

# Close when done
memory.close()

---

Core Concepts

How Activation Works

Every memory has an activation level that determines how easily it can be retrieved:

A = B + S + P + noise

• B (Base-level): Memories accessed recently and frequently have higher B. Follows a power-law decay — rapid drop at first, then slower.
• S (Spreading): If related concepts are active in the current context, connected memories get a boost.
• P (Partial matching): Similar but imperfect matches get adjusted.
• noise: Random variation, like real human recall.

The key insight: A memory accessed 10 times this week has much higher activation than one accessed once a year ago. This is computed mathematically, not heuristically.

Memory Lifecycle

Store --> Active (activation = 1.0)
  |
Access --> Activation increases (access recorded)
  |
Time passes --> Decay reduces activation (every 30 min)
  |
Below threshold --> Decayed (no longer retrievable)
  |
Consolidation --> Patterns extracted as semantic knowledge

---

API Reference

Initialize

from human_memory import Memory, AsyncMemory

# Sync client
memory = Memory(
    base_url="http://localhost:8000",  # Server URL
    timeout=30.0,                      # Request timeout in seconds
)

# Async client (same API, all methods are async)
memory = AsyncMemory("http://localhost:8000")

Context Manager

# Sync -- auto-closes when done
with Memory("http://localhost:8000") as memory:
    memory.store("auto-managed connection")

# Async
async with AsyncMemory("http://localhost:8000") as memory:
    await memory.store("async auto-managed")

---

Storing Memories

result = memory.store(
    content="User asked about Python performance optimization",
    memory_type="episodic",   # Optional: "episodic" (default), "semantic", etc.
    context="support_chat",   # Optional: context tag for the access record
)

Returns: MemoryCreated

Field	Type	Description
memory_id	str	UUID of the stored memory
activation	float	Initial activation (1.0)
created_at	str	ISO timestamp

---

Searching Memories

results = memory.search(
    query="Python performance",  # Natural language query
    top_k=5,                     # Max results (default: 7)
    min_activation=-0.5,         # Optional: override retrieval threshold
)

for r in results:
    print(f"{r.content}")
    print(f"  activation: {r.activation:.3f}")  # ACT-R activation score
    print(f"  similarity: {r.similarity:.3f}")   # Cosine similarity to query
    print(f"  accesses: {r.access_count}")

Returns: list[SearchResult]

Field	Type	Description
memory_id	str	UUID
content	str	Memory text
activation	float	ACT-R activation (ranking key)
similarity	float	Cosine similarity to query
last_accessed	str	When last retrieved
access_count	int	Total access count

Important: Results are ranked by activation (recency + frequency + context), not just similarity. This is what makes the system cognitively plausible.

---

Recalling by ID

info = memory.recall("a1b2c3d4-...")
print(info.content)       # "User prefers dark mode..."
print(info.access_count)  # 6 (incremented by this recall)

Note: Every recall() records an access event, which strengthens the memory. This is the ACT-R reinforcement loop -- useful memories get stronger through use.

Returns: MemoryInfo

Field	Type	Description
memory_id	str	UUID
content	str	Full text
activation	float	Current activation
memory_type	str	"episodic", "semantic", etc.
created_at	str	Creation timestamp
last_accessed	str	Last access timestamp
access_count	int	Total accesses (including this one)

---

Inspecting Full Metadata

detail = memory.inspect("a1b2c3d4-...")

print(f"Activation: {detail.activation:.4f}")
print(f"Salience: {detail.salience:.4f}")
print(f"Emotion: valence={detail.emotion_valence:.2f}, arousal={detail.emotion_arousal:.2f}")
print(f"Decay rate: {detail.decay_rate}")
print(f"Status: {detail.status}")  # "active", "decayed", or "deleted"
print(f"Access history: {len(detail.access_history)} entries")

for access in detail.access_history[:5]:
    print(f"  {access.accessed_at} -- {access.context}")

Returns: MemoryDetail

Field	Type	Description
activation	float	Current ACT-R activation
salience	float	Emotional salience (0-1)
emotion_valence	float	Positive/negative (-1 to 1)
emotion_arousal	float	Calm/excited (0 to 1)
decay_rate	float	ACT-R d parameter (default 0.5)
status	str	"active", "decayed", "deleted"
access_history	list[AccessRecord]	Timestamped access log

---

Forgetting

# Soft-delete a specific memory
deleted = memory.forget("a1b2c3d4-...")  # Returns True

# The memory is now inaccessible
memory.recall("a1b2c3d4-...")  # Raises MemoryNotFound

---

Listing All Memories

# Paginated list (includes active, decayed, and deleted)
items = memory.list(offset=0, limit=50)

for item in items:
    print(f"[{item.status}] {item.content[:60]}... (activation: {item.activation:.2f})")

Returns: list[MemoryItem]

---

Knowledge Graph

The knowledge graph stores semantic memory -- general facts and relationships between concepts. It powers the spreading activation component of ACT-R.

Building the Graph

# Create concepts
memory.add_concept("Python", type="language", activation=0.9)
memory.add_concept("AI", type="field")
memory.add_concept("FastAPI", type="framework")
memory.add_concept("Web Development", type="field")

# Create typed relationships with weights
memory.add_relation("Python", "AI", "USED_FOR", weight=0.9)
memory.add_relation("FastAPI", "Python", "WORKS_WITH", weight=0.95)
memory.add_relation("FastAPI", "Web Development", "USED_FOR", weight=0.85)

Querying Concepts

concept = memory.get_concept("Python")
print(f"{concept.name} ({concept.type})")
print(f"Activation: {concept.activation}")

for rel in concept.relationships:
    print(f"  --{rel.relation_type}--> {rel.target} (weight: {rel.weight})")

Spreading Activation

Find related concepts by spreading activation through the graph:

# What's related to Python?
related = memory.spread(
    concepts=["Python"],  # Active concepts in current context
    depth=2,              # How many hops to traverse (1-4)
    limit=10,             # Max results
)

for r in related:
    print(f"  {r.name}: activation={r.activation:.2f}, spread={r.path_weight:.3f}")

# Output:
#   AI: activation=0.80, spread=0.900
#   FastAPI: activation=0.80, spread=0.950
#   Web Development: activation=0.80, spread=0.808

How it works: When "Python" is the active context, activation spreads through the graph edges. AI gets a boost because Python->AI has weight 0.9. Web Development gets a smaller boost because it's 2 hops away (Python->FastAPI->Web Dev).

Returns: list[SpreadResult]

Field	Type	Description
name	str	Concept name
activation	float	Concept's base activation
path_weight	float	Accumulated spread weight

---

Operations

Temporal Decay

Recalculates activation for all active memories using the ACT-R equation. Memories below the retrieval threshold are marked as decayed.

report = memory.decay()
print(f"Processed: {report.memories_processed}")
print(f"Decayed: {report.memories_decayed}")

Note: Runs automatically every 30 minutes. Use this for manual triggering.

Consolidation

Runs the 4-phase consolidation cycle (inspired by sleep):

1. Replay -- boost high-salience recent memories
2. Extract -- cluster similar episodes, LLM extracts semantic facts
3. Prune -- global activation downscaling, archive weak memories
4. Compile -- convert repeated patterns to procedural skills

report = memory.consolidate()
print(f"Replayed: {report.episodes_replayed}")
print(f"Facts extracted: {report.facts_extracted}")
print(f"Pruned: {report.memories_pruned}")
print(f"Downscaled: {report.memories_downscaled}")

Note: Runs automatically every 6 hours. Use this for manual triggering.

Strategic Forgetting

# Weaken memories not relevant to current goals
report = memory.forget_strategy(
    "strategic_prune",
    goals=["Python", "machine learning", "optimization"]
)
print(f"Memories affected: {report.memories_affected}")

# Archive weakest memories when store is full
report = memory.forget_strategy("capacity_overflow")

Strategy	Description
strategic_prune	Weaken memories unrelated to specified goals
capacity_overflow	Archive weakest memories when capacity exceeded

---

System Status

# Statistics
stats = memory.stats()
print(f"Total: {stats.total}")
print(f"Active: {stats.active}")
print(f"Decayed: {stats.decayed}")
print(f"Deleted: {stats.deleted}")
print(f"Avg activation: {stats.avg_activation:.3f}")

# Backend readiness
checks = memory.ready()
# {'postgres': 'ok', 'qdrant': 'ok', 'neo4j': 'ok'}

---

Async Client

AsyncMemory has the exact same API -- all methods are async:

import asyncio
from human_memory import AsyncMemory

async def main():
    async with AsyncMemory("http://localhost:8000") as memory:
        result = await memory.store("Async memory example")
        results = await memory.search("async example")
        await memory.add_concept("AsyncIO", type="concept")
        stats = await memory.stats()
        print(f"Total memories: {stats.total}")

asyncio.run(main())

Use with FastAPI

from fastapi import FastAPI
from human_memory import AsyncMemory

app = FastAPI()
memory = AsyncMemory("http://memory-server:8000")

@app.post("/chat")
async def chat(message: str):
    # Retrieve relevant context
    context = await memory.search(message, top_k=5)

    # ... call LLM with context ...

    # Store the interaction
    await memory.store(f"User: {message}\nAssistant: {response}")
    return {"response": response}

---

Error Handling

from human_memory import Memory
from human_memory.exceptions import (
    MemoryNotFound,
    ServiceUnavailable,
    ValidationError,
    HumanMemoryError,
)

memory = Memory("http://localhost:8000")

try:
    memory.recall("nonexistent-id")
except MemoryNotFound:
    print("Memory not found")  # 404

try:
    memory.store("")  # Empty content
except ValidationError:
    print("Invalid input")  # 422

try:
    memory.spread(["Python"])  # Neo4j down
except ServiceUnavailable:
    print("Knowledge graph not available")  # 503

Exception hierarchy:

HumanMemoryError          # Base (any HTTP error)
  MemoryNotFound          # 404
  ServiceUnavailable      # 503
  ValidationError         # 422

---

Complete Example: AI Agent with Memory

from human_memory import Memory

def create_agent():
    memory = Memory("http://localhost:8000")

    # Build the agent's knowledge base
    memory.add_concept("Python", type="language")
    memory.add_concept("JavaScript", type="language")
    memory.add_concept("Web", type="field")
    memory.add_relation("JavaScript", "Web", "USED_FOR", weight=0.95)
    memory.add_relation("Python", "Web", "USED_FOR", weight=0.6)

    return memory

def agent_respond(memory: Memory, user_message: str) -> str:
    # 1. Retrieve relevant memories
    memories = memory.search(user_message, top_k=5)
    context = "\n".join(
        f"- {m.content} (relevance: {m.activation:.2f})"
        for m in memories
    )

    # 2. Check knowledge graph for related concepts
    try:
        related = memory.spread(["Python"], limit=3)
        for r in related:
            context += f"\n- Related: {r.name} (weight: {r.path_weight:.2f})"
    except Exception:
        pass

    # 3. Build prompt for LLM
    prompt = f"""You are a helpful assistant with memory.

Relevant context from past interactions:
{context if context else "No relevant memories yet."}

User: {user_message}
Assistant:"""

    # 4. Call your LLM here
    response = call_your_llm(prompt)

    # 5. Store this interaction
    memory.store(f"User asked: {user_message}")

    return response

# Usage
memory = create_agent()
print(agent_respond(memory, "How do I optimize Python code?"))
print(agent_respond(memory, "What about JavaScript for web apps?"))

# Maintenance (also runs automatically in background)
memory.decay()
memory.consolidate()

stats = memory.stats()
print(f"Agent has {stats.active} active memories")
memory.close()

---

All Models

Model	Returned by	Key fields
MemoryCreated	store()	memory_id, activation, created_at
SearchResult	search()	memory_id, content, activation, similarity, access_count
MemoryInfo	recall()	memory_id, content, activation, memory_type, access_count
MemoryDetail	inspect()	All of MemoryInfo + salience, emotion_valence, emotion_arousal, decay_rate, status, access_history
MemoryItem	list()	memory_id, content, activation, salience, status, access_count
Stats	stats()	total, active, decayed, deleted, avg_activation
Concept	add_concept(), get_concept()	name, type, activation, relationships
Relation	add_relation()	source, target, relation_type, weight
SpreadResult	spread()	name, activation, path_weight
DecayReport	decay()	memories_processed, memories_decayed
ConsolidationReport	consolidate()	episodes_replayed, facts_extracted, memories_pruned, memories_downscaled
ForgetReport	forget_strategy()	strategy, memories_affected
AccessRecord	Inside MemoryDetail	accessed_at, context

---

License

MIT