decision · choice
The events log uses PostgreSQL
Confidence
0.8 → 0.82
Evidence
benchmark_result_id
The events log moves to MongoDB
Reason
New benchmarks show MongoDB handles our access patterns better.
SmartMemory keeps your agent's decisions and lessons as typed records, with as-of recall and a tamper-evident audit trail.
YouSet up the billing tables in MongoDB.
Before I do: your team decided billing uses a relational database, because ledger writes need relational guarantees. A document store was ruled out. Shall I use PostgreSQL instead?
Which database does billing use?
A relational database.
Illustrative team memory and dates.
Built for trust
A knowledge graph says what exists. A context graph also keeps when it was true, where it came from, why it was recalled, and who may see it.
Every fact records when it was true and when SmartMemory learned it. Old facts are superseded, not deleted, so you can ask what was believed on any past date.
Every memory records where it came from, and decisions keep the evidence behind them, so you can trace a claim back to its source.
Memory changes are kept in a tamper-evident, hash-chained history, and an explain call shows how a memory came to be.
Every read and write is scoped to a workspace, user, or tenant, so one team never sees another team’s memory.
See it in action
A real terminal session: install SmartMemory, seed a project, ask it questions, and watch a brand-new session pick up the context at a fraction of the tokens. Under six minutes.
More on YouTube: Introduction · Getting started (all 8 parts) · SmartMemory in action
Decision memory
SmartMemory records what your agent decided. A decision can carry the reason behind it and a confidence score. Supporting evidence raises that score. When a decision is superseded or retracted, it stays in the record with the reason.
decision · choice
Confidence
0.8 → 0.82
Evidence
benchmark_result_id
The events log moves to MongoDB
Reason
New benchmarks show MongoDB handles our access patterns better.
Why SmartMemory
Keep each choice with its reason and the evidence behind it, as one typed record.
Ask with an as-of date to recover the decision that was current then. On 25 synthetic reversed-decision questions it answered all 25 correctly on two seeds and two models. A notes file edited in place got 1 to 3.
Follow relationships among decisions, constraints, lessons, and reasoning instead of searching interchangeable notes.
In our test at 500 and 2,000 facts, SmartMemory injected about 470 tokens where a curated markdown file took about 11,700 to 12,000, roughly 25 times less, with no loss in answer quality. Single seed, 28 questions per size, synthetic coding-project corpus.
How teams lose what they already knew, one use case each.
Everything you need to know
Give every agent the judgment, constraints, and history it needs to move forward.