// SUBSTRATE · v0.1.0
session-loam
Memory continuity for autonomous agents.
A SQLite-backed substrate that remembers what your agent learned across runs. Identity-scoped. Tag + full-text retrieval. Never-delete with reinforce-on-retrieve. No daemon, no broker, no service to operate — just a file per agent and a Python library that knows how to query it.
// THE FAILURE MODE
"I solved this same problem three sessions ago." But the agent has no durable record of the solution, so it solves it again — possibly worse. Chat history compresses away. Vector DBs lose identity. Ad-hoc files lose schema. Learning-as-state has no substrate. Until now.
Three problems, named
Chat-history-as-memory
The conversation is volatile. Compaction strips the context that gave answers their shape. Nothing accumulates between runs.
Vector-DB-as-memory
ChromaDB, Pinecone, mem0. Great for semantic similarity. No identity scoping, no tag filters, no time queries, no attestation pathway.
Notes-folder-as-memory
A pile of markdown files. No schema, no retrieval, no reinforcement signal. The agent can't find what it wrote last week.
Three primitives
PRIMITIVE 1
One file per agent
~/.session-loam/<agent>.db.
SQLite, ACID, queryable by any client. Identity is the partition; agents don't
accidentally cross-pollinate.
PRIMITIVE 2
FTS5 + tags + time
Full-text search via SQLite's FTS5 virtual table. Tags via a junction table for exact-match filtering. Time-range queries via indexed timestamps. No embedding model in the critical path.
PRIMITIVE 3
Reinforce on retrieve
Every hit bumps last_accessed +
access_count. Memories that matter
surface to the top of ranked search. Memories that don't fade quietly — until you prune them.
// THE NEVER-DELETE RULE
Memories are not silently expired. They persist until you explicitly prune them. Edits are supersedes — the new entry references the old, the old stays for audit.
Search hides superseded entries by default and uses bm25 + recency + confidence + access-count to rank what surfaces. Tunable per call.
Where it slots in
session-loam is one library you can drop into a Python or shell pipeline. It composes with three substrate-layer projects you might already know, and works fine without any of them.
USE-CASE
Day-to-day patterns
Compaction bridge. Per-task memory recall. Cross-session preferences. User-specific facts. The five concrete shapes of work where session-loam saves real time.
EXAMPLES
Wired up to other substrates
PreCompact bridge to session-essence.
Memory-aware worker handlers via swarm-lib.
Attestation flow with tribunal (v0.2).
DOCS
Concepts
Entry schema, identity scoping, supersede semantics, retrieval ranking, FTS5 query shape, concurrent-writer model.
DOCS
vs. mem0 / ChromaDB / MCP
When the existing memory tools are the right answer, and when they aren't. Honest comparison.
60-second tour
# Install (once published)
pip install session-loam
# Open the store for the current agent identity
python -c "
from session_loam import Store, Entry
store = Store.open(agent='mabus')
store.write(Entry(
type='observation',
content='Dan prefers terse responses with no trailing summaries.',
tags=['dan', 'preference'],
confidence=0.95,
))
results = store.search(query='terse responses', tags=['preference'])
for r in results:
print(r.entry.content)
""Identity is stable. Work-state is ephemeral. There's no canonical substrate for the middle thing — what an agent has learned across runs. session-loam is the substrate-layer answer: filesystem-backed, identity-aware, never-delete by default, ready to compose."