agent-memory-store

Lightweight persistent memory for AI agents. Drop in, zero dependencies, privacy-first.

SQLite backend. Four memory types. TTL expiry. Full audit log. Built-in forget() for GDPR compliance. Override _score_relevance() to plug in vector search.

---

Why not MemOS / MemU / A-MEM?

Every existing agent memory framework has the same problems: complex setup (Docker, Postgres, pgvector, embedding models), no privacy controls, no audit log, and they want to own your entire agent architecture.

agent-memory-store is different:

	MemOS / MemU	agent-memory-store
Setup	Docker + Postgres + pgvector	pip install agent-memory-store
Dependencies	20+	0 mandatory
Backend	Postgres with pgvector	SQLite (stdlib)
forget(user_id=...)	✗	✓
Audit log	✗	✓
Drop into existing agent	Requires rewrite	3 lines
Vector search	Built-in	Override one method

---

Install

pip install agent-memory-store   # zero mandatory dependencies

---

Quick Start

from agent_memory import MemoryStore

mem = MemoryStore("agent.db")

# Store memories
mem.add("user prefers concise responses", tags=["preference"], importance=0.8)
mem.add("user works in NHS radiology", memory_type=MemoryType.SEMANTIC, importance=0.9)
mem.add("last session: discussed Python async", memory_type=MemoryType.EPISODIC)

# Get context for a prompt — ready to inject
context = mem.context_for("communication style")
system_prompt = f"You are a helpful assistant.\n\nWhat I know:\n{context}"

# Search
results = mem.search("response format", limit=3)
for r in results:
    print(f"{r.relevance_score:.2f}  {r.content}")

# Privacy: forget everything for a user
mem.forget(user_id="user_123")          # right to erasure

# Data minimisation
mem.forget(older_than_days=90)          # forget anything older than 90 days
mem.forget(tags=["sensitive"])          # forget by category

# Audit trail
for event in mem.audit_log():
    print(event)

# Stats
print(mem.stats())

---

Memory Types

Modelled on cognitive science memory taxonomy:

from agent_memory import MemoryType

MemoryType.SEMANTIC    # General facts: "user works in healthcare"
MemoryType.EPISODIC    # Specific events: "user asked about X on Monday"
MemoryType.PROCEDURAL  # How to do things: "user prefers step-by-step responses"
MemoryType.WORKING     # Temporary context for the current task (set ttl_days=1)

---

Privacy Controls

Built for GDPR Article 17 (Right to Erasure) and UK GDPR:

# Forget a specific user — right to erasure
deleted = mem.forget(user_id="user_123")
print(f"Deleted {deleted} memories")

# Forget everything for an agent
mem.forget(agent_id="my_assistant")

# Data minimisation — delete old memories
mem.forget(older_than_days=30)

# Category-based deletion
mem.forget(tags=["sensitive", "pii"])

# Auto-expire working memory
mem.add("temporary task context", memory_type=MemoryType.WORKING, ttl_days=1)
mem.expire_stale()  # call at session start

Every forget() call is written to the audit log.

---

Audit Log

Every operation creates an immutable audit event:

for event in mem.audit_log(limit=20):
    print(event)
# AuditEvent(2026-03-28 14:30:01 | add    | agent=assistant mem=mem_abc123 | type=semantic tags=['preference'])
# AuditEvent(2026-03-28 14:30:05 | search | agent=assistant | query='response style' hits=2)
# AuditEvent(2026-03-28 14:35:12 | forget | agent=assistant | user_id=user_123 deleted=4)

# Filter by agent or action
events = mem.audit_log(agent_id="assistant", action=AuditAction.FORGET)

---

Plug in Vector Search

The default scorer uses keyword overlap. Override one method for semantic search:

from agent_memory import MemoryStore, Memory

class SemanticMemoryStore(MemoryStore):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        from sentence_transformers import SentenceTransformer
        self.model = SentenceTransformer("all-MiniLM-L6-v2")

    def _score_relevance(self, query: str, memory: Memory) -> float:
        import numpy as np
        q = self.model.encode(query)
        m = self.model.encode(memory.content)
        cos_sim = float(np.dot(q, m) / (np.linalg.norm(q) * np.linalg.norm(m)))
        return cos_sim * 0.7 + memory.importance * 0.3

mem = SemanticMemoryStore("agent.db")
# Everything else works exactly the same

---

Multi-Agent & Multi-User

# Same database, different agents
mem = MemoryStore("shared.db")

mem.add("prefers formal tone", agent_id="sales_agent", user_id="user_001")
mem.add("technical background", agent_id="support_agent", user_id="user_001")

# Search is scoped by agent
sales_context = mem.context_for("tone", agent_id="sales_agent")
support_context = mem.context_for("background", agent_id="support_agent")

---

CLI

agent-memory stats agent.db
agent-memory list agent.db --agent my_assistant
agent-memory search agent.db "user preferences"
agent-memory audit agent.db --limit 50
agent-memory forget agent.db --older-than 90
agent-memory export agent.db backup.json

---

Related Projects

• llm-extract — Extract structured data from any document using LLMs
• agentic-alignment-toolkit — Evaluate goal-alignment and oversight gaps in autonomous AI pipelines

---

Linda Oraegbunam | LinkedIn | Twitter