Long-term memory in Logicore is powered by AgentrySimpleMem and stored in a LanceDB vector database. Unlike short-term session history, it persists across process restarts and can optionally be shared across multiple sessions.
Enable it by passing memory=True to any agent constructor:
from logicore.agents.agent import Agent
agent = Agent(llm="ollama", memory=True)
What Gets Stored
Not every message turn is stored. AgentrySimpleMem applies a scoring pipeline to filter out low-signal content before it ever reaches LanceDB.
Signal scoring
Each dialogue turn is scored using _score_memory_signal(). Points are added or subtracted based on pattern matching:
| Signal | Score change |
|---|
Contains durable markers (my name is, I prefer, stack, framework, deadline, etc.) | +2 |
References a completed action (scheduled, created, saved, updated, fixed) | +2 |
| Contains a number or version | +1 |
| Message length in the 5–40 token sweet spot | +1 |
| Is a question | -2 |
Is vague acknowledgement (ok, thanks, sure, hi) | -2 |
Is a transient reminder request (remind me in 5 minutes) | -3 |
Is a promissory assistant statement (I'll remind you) with no confirmation | -2 |
SIMPLEMEM_MIN_STORE_SCORE (default: 3) proceed to fact extraction.
Adjust the threshold with the SIMPLEMEM_MIN_STORE_SCORE environment variable. Lower values store more facts; higher values are more selective.
_extract_atomic_facts(). Each sentence or bullet-point line is evaluated independently. Facts are capped at SIMPLEMEM_MAX_MEMORY_CHARS characters (default: 220) and at SIMPLEMEM_MAX_FACTS_PER_TURN facts per turn (default: 3).
Storage Pipeline
MemoryEntry schema
Each stored entry is a MemoryEntry dataclass with the following fields:
@dataclass
class MemoryEntry:
entry_id: str # UUID
lossless_restatement: str # The stored fact text (with [Speaker][score=N] prefix)
keywords: List[str] # Extracted keywords (stop-words removed)
timestamp: Optional[str] # ISO 8601 datetime of when it was queued
location: Optional[str]
persons: List[str] # Speaker name
entities: List[str]
topic: Optional[str]
lossless_restatement field carries a score tag, for example:
[User][score=4] My preferred stack is FastAPI + PostgreSQL
SIMPLEMEM_MIN_RETRIEVE_SCORE threshold.
Retrieval Pipeline
Retrieval happens at the start of every agent.chat() call when memory is enabled.
AgentrySimpleMem.on_user_message() handles this and returns a list of relevant memory strings:
contexts: List[str] = await agent.simplemem.on_user_message("What stack do I use?")
# e.g. ["[User][score=4] My preferred stack is FastAPI + PostgreSQL"]
Retrieval threshold
Only results whose embedded score tag meets SIMPLEMEM_MIN_RETRIEVE_SCORE (default: 2) are returned. This prevents low-confidence facts from polluting the context.
Persistence Scope
Long-term memory can be scoped in two ways, controlled by the isolate_by_session parameter on AgentrySimpleMem.
Each session_id gets its own LanceDB table:memories_<user_id>_<session_id>
isolate_by_session=True). Facts written in one session are not visible in another session, even for the same user.# Default behavior — session-isolated tables
agent = Agent(llm="ollama", memory=True)
await agent.chat("I use PostgreSQL", session_id="session-1")
# session-2 cannot retrieve the PostgreSQL fact
reply = await agent.chat("What DB do I use?", session_id="session-2")
To share facts across sessions for the same user, override the simplemem instance after agent creation:from logicore.agents.agent import Agent
from logicore.simplemem import AgentrySimpleMem
agent = Agent(llm="ollama", memory=True)
agent.simplemem = AgentrySimpleMem(
user_id=agent.role,
session_id="global",
isolate_by_session=False, # all sessions share one table: memories_<user>
debug=True
)
# Both sessions read from and write to the same table
await agent.chat("I prefer dark mode", session_id="session-1")
reply = await agent.chat("What do I prefer?", session_id="session-2")
print(reply) # Can recall the dark mode preference
Storage Location
LanceDB data is stored locally by default under:
logicore/user_data/lancedb_data/
LANCEDB_PATH to an S3 or Azure Blob path and set AGENTRY_MODE=cloud.
Environment Variables
| Variable | Default | Description |
|---|
AGENTRY_MODE | local | Set to cloud to use LANCEDB_PATH for remote storage |
LANCEDB_PATH | (derived) | Override LanceDB storage path in cloud mode |
EMBEDDING_MODEL | qwen3-embedding:0.6b | Ollama embedding model name |
OLLAMA_URL | http://localhost:11434 | Ollama service URL |
SIMPLEMEM_MIN_STORE_SCORE | 3 | Minimum signal score to persist a fact |
SIMPLEMEM_MIN_RETRIEVE_SCORE | 2 | Minimum score tag to include a fact at retrieval |
SIMPLEMEM_MAX_FACTS_PER_TURN | 3 | Max atomic facts extracted per dialogue turn |
SIMPLEMEM_MAX_MEMORY_CHARS | 220 | Max character length for a single stored fact |
SIMPLEMEM_TOP_K | 5 | Number of results returned by semantic search |
Memory Flushing
The dialogue queue is processed automatically:
- After the assistant’s final response in each
agent.chat() call
- During agent cleanup
To flush manually:
if agent.simplemem:
await agent.simplemem.process_pending()
If the process is killed before process_pending() runs, any queued but unprocessed dialogue turns will be lost. For critical workloads, call process_pending() explicitly after important turns.
Inspecting Memory
# View memory statistics
if agent.simplemem:
stats = agent.simplemem.get_stats()
print(stats)
# {
# "user_id": "alice",
# "session_id": "sprint",
# "table_name": "memories_alice_sprint",
# "initialized": True,
# "pending_dialogues": 0,
# "total_memories": 12
# }
# Wipe all stored memories for this user/session
agent.simplemem.clear_memories()