- CapturedA design note records the two problems left open after the first version: verification runs permissively by default, so enforcing it has wide blast radius, and a revocation can race a concurrent change.
- Cited 2 days later, and again on day 65Five separate sessions pulled it back, eleven citations in all.
- EffectLater design work started from the known-open problems instead of rediscovering them.
Shared memory for teams that build with agents
Every assertion compounds what your company knows.
Assertion turns what your teams decide, build and learn into knowledge the company keeps and AI reuses.
Staying on Claude Code, Cursor, or Codex? The free plugin brings the same memory there.
15,988 countingclaims captured so far, each with the evidence behind it
The problem
Every session starts from zero.
The constraint was settled three weeks ago, in someone else’s terminal. Today’s session cannot see it, so the work asks again — or quietly decides otherwise.
01
The same decision, re-argued
A constraint settled in March is debated again in May, because nothing carried it forward and nobody thought to look.
02
Reasoning that dies with the window
Why it was done that way lives in one person’s scrollback. When the session closes, the answer goes with it.
03
Re-explaining, every time
Each new session opens by restating what the team already established, to an assistant that was there for none of it.
What happens instead
The next session already knows.
A decision made once is captured with the reasoning behind it. Weeks later, in a session that was not there, the work starts from it and says where it came from.
See it in customer sessions
Example, not a customer’s data
payments-apiLast Tuesday
New session
payments-apiToday
The point is not to preserve every word. It is to make the settled decision and its rationale available when they become relevant again.
Early evidence
What recalled context actually changed.
Based on early customer data and internal benchmarks, measured conservatively, with every definition published.
What compounding looks like
Three sequences from customers’ memory.
In each one the session that produced the knowledge had ended long before the work that used it began, and nobody went looking for it. Every date and count below is real. The subject matter has been changed.
- CapturedA slowdown is diagnosed as a scale-and-deadline problem, explicitly not a deadlock and not a correctness defect. The job runs through thousands of iterations and simply never returns in time.
- Cited across the next 23 daysFour sessions, six citations.
- EffectNobody reopened the deadlock theory the evidence had already ruled out.
- CapturedAn intermittent build failure is written off as environmental, since the same commit passed elsewhere minutes earlier.
- CorrectedIt turns out to be a real production race. The assertion is updated to say so, and the fix is recorded against it.
- Cited 23 days laterThe later session received the correction, not the original wrong call.
Of the exchanges where recalled context materially helped, roughly two-thirds used a fact captured in a different session. An exchange is one question and the work that answers it.
Estimated time saved per ten exchanges in which memory surfaced, mostly from not restating or re-researching.
Cheaper to run than an agent carrying its whole history, at equal factual recall, once a session runs long enough for that history to weigh. The longer it runs, the wider the gap.
How to read these figures
The three sequences are real events from customers’ memory, retold with the subject matter changed; the dates, session counts and citation counts are unaltered. The card figures come from the same sessions, where a recall counts as helped only when the agent cited an assertion. The cost figure is a range from internal benchmarks at current model rates, not a single measurement: it grows with session length, because a full-history agent re-sends a transcript that keeps growing while memory sends the same distilled set. We will publish the benchmark.
Trust by construction
See what is remembered, and replace what is wrong.
Assertion is designed around explicit claims and evidence, so you can see what is remembered, where it came from, and what replaced it.
Distilled memory, not a repository copy
The memory layer stores distilled claims and short evidence excerpts. Each exchange is processed to produce them, then discarded; complete repositories, uploaded files, and full transcripts are never stored as memory.
You choose who can see it
What Assertion remembers stays private to you until you share it with a team or a project, and then only with the people in it.
Your work is not training data
Your repository content, uploaded data, and assertions are not used to train models.
Corrections leave a history
Evidence remains attached to the claim. When a claim becomes wrong, its replacement supersedes it without silently erasing the previous decision.
Ways to use it
Use our agent, or keep the one you have.
The same memory either way. The difference is which agent runs it.
See pricingPricing
Free to start, either way.
Use Assertion’s own agent, or add the free plugin to yours.
Assertion MemoryThe agent is the product. The plugin brings the same memory to the tool you already use, free.
Long sessions that do not degrade.
Free to start
- Builderthe baseline$20 monthly · $16 annually
- Pro3× the usage$60 monthly · $48 annually
- Max10× the usage$200 monthly · $160 annually
Teams, per member, usage pooledTeam $40 monthly or $32 annually. Team Premium $80 monthly or $64 annually, at twice a Team seat's usage, with admin controls.
- Context stays flat as a session runs, so early work is never dropped mid-task
- The same frontier models, at the same price, with more work included
- Memory native from the first prompt, not injected into someone else’s loop
- A free tier to start, with weekly usage included
or in a terminal: curl -sSL https://storage.googleapis.com/assertion-downloads/install.sh | sh
Keep your agent. Add shared memory.
Free for personal use
Team Memory, per member$10 monthly or $8 annually. 14-day trial, no seat minimum.
- Claude Code, Cursor, and Codex
- Personal memory is free, however much you capture
- Shared spaces: teammates recall each other’s work
- Same client, same commands, nothing to relearn
The plugin stays free for personal use, for anyone who would rather keep the agent they have. Annual billing saves 20% on every paid plan. Agent tiers differ by included usage, not capability. Team Memory is included with native-agent team seats. Assertion Analytics is licensed annually and includes Memory.
Questions
The short answers.
What exactly does Assertion store?
Distilled claims, supporting evidence excerpts, and the relationships needed to retrieve and revise them. Assertion does not use complete transcripts as memory.
Who can see my memory?
Only you and members of the spaces with which you explicitly share it.
What happens when a decision changes?
The new assertion supersedes the old one. The previous decision remains available as history but is no longer presented as current guidance.
How does this relate to Assertion Analytics?
Analytics is our second product, for business-data analysis rather than software work. It is built on the same memory and includes it. If that is what you need, start there instead.
Where do the numbers come from?
Early customer sessions and internal benchmarks, measured conservatively; the definitions are stated next to each figure. More data and a reproducible benchmark will be published as the sample grows.
Is my code or data used to train models?
No. Your content is never used to train models.
The next step for AI
Start from what your company already knows.
Memory carries decisions, evidence and context across sessions and teammates, so each piece of work starts from what the team already established. We think this is the stage after agentic workflows, and the road to human-level reasoning in AI.