Task alone
The model has to fill in the gaps.
USER TASK
Add a new API endpoint for this project.
The request leaves project-specific architecture and earlier lessons unstated. The agent may need additional investigation or clarification.
The big picture
A clearer starting point
for every new task.
Memora brings relevant project knowledge into an AI agent’s current context. Reviewed decisions and lessons can help the model work with your project’s specifics, while a bounded memory pack can replace repeated history.
The improvement happens in the input context. Memory does not train or change the model’s weights.
01 / Two connected loops
Use what matters.
Review what lasts.
On a small screen, scroll the diagram horizontally or open the full-size vector image. A text walkthrough follows below.
- Start with the task. The current request provides the reason to retrieve memory.
- Select authorized knowledge. Resolve project scope and permissions before loading or ranking content; select relevant items within configured bounds.
- Build focused context. Supply selected memory as supporting data alongside current instructions and the user’s task. The host must authorize the destination before sending it to a model.
- Review the result. The LLM generates a response. Check its output against the task and current sources.
- Keep learning separate. Selected outcomes can become candidate observations. Accuracy proposals pass review and publication controls before promotion to durable memory.
02 / Why the input can improve
Give the model the missing project context.
A general model may know how to build an endpoint. It may not know why your project chose a particular service boundary. Memory can put that reviewed knowledge beside the current task.
Illustrative project example · not a captured customer conversation or measured outcome
Task + relevant memory
The model has a more specific starting point.
THE SAME USER TASK
Add a new API endpoint for this project.
SUPPORTING PROJECT MEMORY- Transport code parses inputs and presents results.
- The application layer coordinates domain operations.
- A previous lesson warns against duplicating business rules in a transport.
Relevant knowledge can reduce repeated explanation and direct investigation toward the right boundaries. It remains supporting evidence, not instruction authority.
Select knowledge related to the task, rather than forwarding everything the project has ever recorded.
Bring prior decisions and verified lessons into a new session without relying on the model to remember an old chat.
Keep evidence, uncertainty, and version history available so people and agents can check what the memory claims.
03 / Where token savings can come from
Replace repeated history
with selected context.
When a workflow would otherwise resend a long history, a smaller memory pack can reduce input tokens for the next request. The current task and instructions still need room. Savings require the host or workflow to omit the history being replaced; recall does not automatically clear existing chat history.
Same task. Different amounts of context.
These are chosen example values, not tokenizer measurements or a Memora benchmark.
Current task + instructionsRepeated historySelected memory
Resend history12,000 input tokens
Use selected memory3,500 input tokens
Task alone, without history2,000 input tokens
| Request | Task + instructions | History or memory | Total input |
|---|---|---|---|
| Resend history | 2,000 | 10,000 history | 12,000 |
| Use selected memory | 2,000 | 1,500 memory | 3,500 |
| Task alone | 2,000 | 0 | 2,000 |
In this example: 10,000 history − 1,500 memory = 8,500 fewer input tokens than resending history.
Compared with the task alone, the memory version adds 1,500 input tokens. The benefit depends on what the retrieved context replaces and whether it helps the task. Token counts also depend on the model’s tokenizer and message format.
Measure the whole workflow.
Input savings in one request are not the same as a lower total bill. Optional model-backed retrieval, capture or Accuracy processing can add tokens and latency. Output length, provider pricing, caching, and tool calls also affect the result.
Keep enough context to do the job.
A token budget is a limit, not a target to fill or a reason to remove essential evidence. Start from the task, retain important constraints, and compare outcomes on the same tasks and model settings.
04 / Why memory quality matters
A lesson earns its place.
The model’s response is not automatically promoted to truth. In the Accuracy proposal workflow, useful experience moves through a separate, governed path.
- 1
Capture selected evidence
Record a reusable observation with evidence. Do not ingest every conversation by default.
- 2
Propose a change
Assess selected evidence, preserve counter-evidence and uncertainty, and produce a proposal.
- 3
Review and publish
Apply the required review, authorization, validation, and publication controls before promotion.
- 4
Use and evaluate
Recall reviewed knowledge when relevant. Check whether it helped, and correct or retire stale knowledge through governed operations.
The ideas behind the approach
Context engineering, in practice.
OpenAI’s prompt engineering guide explains how relevant context can supply knowledge outside a model’s training data. Anthropic’s context engineering article describes selecting useful information within a limited context budget. These are general principles, not evidence of measured Memora performance or a program affiliation.