MemoryEngine · PAI · opencode & Claude Code
Not a vector database. Not a text search engine. An episodic memory.
Positronic is a polytemporal memory engine: it distills every turn into typed events and objects and recalls them in under a millisecond. Local SQLite by default, PostgreSQL when you centralise, and a horizon from days to forever.
npm i -g @positronic-ai/opencode-plugin · AGPL or commercial
Validated on
- LongMemEval
- RULER
- E7 synthetic
- SQLite or PostgreSQL
- S3 Object Lock
- MIT-licensed sure stack
The problem
Agents forget, or they drown.
Most memory systems make one of two bets. They store everything and never forget, so the context window swells until retrieval is a haystack scan. Or they summarize aggressively and quietly delete the detail you needed three weeks later.
Positronic does neither. Every event carries a time vector — wall-clock, arrival order, subjective τ, and a confidence interval — and retention runs on survival curves instead of a TTL. Memory demotes; it is never deleted.
How it works
Recall before you resume
- Record — every turn auto-ingests to
.positronic/brains/<name>/memory.db. No cloud, no manual calls. - Recall — episodes, digests, and dossiers return ranked, each carrying its own τ/wall/kind.
- Resume — the agent continues with context it actually kept.
Why teams pick it
Built to remember, not to hoard.
Polytemporal retention
Retention is a curve, not a counter. Decay runs on subjective time, so old-but-important memories survive and recent noise fades on schedule.
Curation at the source
A salience gate scores surprise × goal-weight at encode time. Not everything gets stored — which is why recall stays cheap.
Tensor-grounded objects
Entities persist across episodes as evolving objects. Activate, reconstruct, decay, fuse — deterministic steps you can audit.
Sub-millisecond recall
SQLite, FTS5, and reciprocal-rank fusion deliver p95 0.7 ms. Top-8 retrieval uses ~242 tokens regardless of haystack size.
Local-first, central on demand
A brain is a plain SQLite file — query it with SELECT *. Point it at PostgreSQL for a centralised team brain instead. Nothing leaves your machine unless you choose.
Federated when you want
Run one brain per concern, then fuse across all of them with RRF — or expose one over HTTP and share it across hosts.
The numbers
Two honest numbers.
The store — p95 0.7 ms on the synthetic harness, ~6k events/s ingest.
The pipeline — recall 1.000 across 500 LongMemEval trials; accuracy 0.262 vs 0.202 without memory. Top-8 retrieval holds a flat ~242 tokens while the full haystack grows from 2.2k to 18k.
The store is instant. The pipeline takes what the LLMs take. Both are real.
Retention
Pick a curve, not a deadline
| Profile | Sbase | Horizon | Kept @ 78 wks |
|---|---|---|---|
| archival | 1e6 | forever | 55 / 55 |
| long_term | 120 | months | 55 / 55 |
| balanced | 30 | weeks | 35 / 55 |
| short_term | 6 | days | 11 / 55 |
Anything worth keeping permanently can go to an S3 Object Lock archive — GOVERNANCE or COMPLIANCE retention. The lock is re-read after every write, so a copy reports archived only once it is provably immutable.
Questions
Good to know
How is this different from a vector database?
A vector database stores embeddings and returns nearest neighbours. Positronic stores typed events on a temporal schema, gates them at encode time, and recalls through deterministic steps — so results are reproducible and every row is inspectable with plain SQL.
Do I need a GPU?
No. Tier-1 lexical recall (SQLite FTS5) runs anywhere. Tier-3 semantic recall adds BGE-M3 on your GPU if you want it.
Does my data leave my machine?
Not by default. A brain starts as a local SQLite file on your machine. Centralising on PostgreSQL, federation, the HTTP server and S3 archival are all opt-in — you choose when memory moves.
What does it cost?
The engine is AGPL-3.0-or-later or commercial. The sure-* UI stack is MIT-licensed. Contact us for a commercial license.
Give your agent a memory it can resume from.
Questions? info@positronic-ai.us