KeepMind's Projetos view stores the state needed to resume work: objective, success criteria, next step, decisions, rejected approaches, artifacts, evidence, observed outcomes and unanswered questions. The native chat and the bundled MCP server use the same continuity engine. No model subscription is needed to create or review a project.
- Open Projetos in the desktop app and create one project with an objective,
success criteria and a next step. Sources are optional, existing Markdown paths
within the current space, such as
notes/research.md. - Record a decision with its alternatives, assumptions and expected result. New decisions start as proposals. Review them in the interface to accept, reject or supersede them; previous reviews remain in the record.
- Record approaches already tried, why they failed or were rejected, and when they might still be useful. Link an artifact to the decisions it depends on.
- Use Copiar contexto to take the current handoff to another assistant, or connect that assistant to the same memory through MCP.
- Record observed results separately from execution status and expected results. A successful command does not automatically confirm a lesson or a business outcome. Confirm or reject an existing proposed lesson in the human interface; its review history is retained.
The standalone Corp backend does not yet implement this API. Its shared interface shows an explicit availability message rather than pretending to save projects.
Keep the server entry point and both modules together:
tools/keepmind-mcp.mjs
tools/lib/memory-contract.mjs
tools/lib/continuity.mjs
The app materializes the complete bundle for external agents. New memory repositories receive the modules through scaffolding. Existing repository files are preserved: when updating a manually copied server, update all three files. The standalone npm package includes the modules; it still needs publication before a registry installation receives a new version.
Encrypted access requires the existing unlocked native key service and explicit launcher scopes. Client write permission must be enabled by the launcher; tool arguments cannot grant it. See SECURITY-MODEL.md and the server README for the existing unlock mechanism.
| Tool | Purpose |
|---|---|
continuity_list |
List authorized projects in the client's default space |
continuity_create |
Create a project and optionally snapshot its source versions |
continuity_get |
Read the full structured project and current revision |
continuity_context |
Obtain categorized context, evidence states and fingerprint |
continuity_update |
Update title, objective, success criteria, next step or status |
continuity_add_decision |
Propose a decision, alternatives and expected result |
continuity_add_attempt |
Preserve an attempted approach, result and applicability |
continuity_add_artifact |
Associate a Markdown artifact with decisions and sources |
continuity_record_outcome |
Record execution status, observed result and proposed lesson |
continuity_add_question / continuity_resolve_question |
Track explicit information gaps |
continuity_check |
Check source versions and dependent artifacts |
continuity_metrics |
Count records and human usefulness feedback |
For example, the arguments to create and resume a project are:
{
"title": "New onboarding",
"objective": "Help a new user resume one real project",
"successCriteria": "The user can identify the next step without repeating context",
"nextStep": "Test the first project flow",
"evidence": ["notes/user-research.md"]
}Call continuity_context with the returned projectId. Then use the returned
revision when proposing a decision:
{
"projectId": "<returned UUID>",
"expectedRevision": 1,
"choice": "Start onboarding with one project",
"alternatives": ["Start with an empty graph"],
"rejectionReason": "An empty graph does not explain the next useful action",
"assumptions": ["The user already has one project to resume"],
"expectedOutcome": "The user can resume work with less repeated explanation",
"evidence": ["notes/user-research.md"]
}All changes to an existing project require expectedRevision. On
REVISION_CONFLICT, read the current state and reconcile before retrying. Do not
blindly retry the previous write or replace the raw project file. The generic
sb_write and sb_append tools refuse protected project records, including path
aliases. A model-supplied actor: "human" cannot authorize a human action.
MCP clients can read analysis reports through continuity_get and their stale
state through continuity_context. They cannot run the native analysis or perform
human reviews. Decision and analysis reviews, lesson confirmation or rejection,
evidence acknowledgement and feedback are reserved for the trusted desktop
interface; the engine rejects direct agent attempts as well.
continuity_context separates accepted decisions, proposals and rejected or
superseded decisions. It preserves attempts, predictions, assumptions, observed
results, open questions and source versions. It includes:
revision: the stored project version used for subsequent writes.fingerprint: SHA-256 of the structured context and current evidence state.evidence: recorded source digest, current digest and availability status.impacts: changed or unavailable sources and artifacts depending on affected decisions.missingQuestion: a deterministic suggestion based on missing objective, success criteria, expected result, source review or next step.
The default response budget is 65,536 UTF-8 bytes, including its readable text.
maxBytes can explicitly request between 1,024 and 1,048,576 bytes. An oversized
response fails with CONTEXT_BUDGET_EXCEEDED; it never silently drops constraints
or decisions. The caller can explicitly increase the budget or inspect individual
project records with continuity_get.
Fingerprints identify a context version; they do not prove a statement true.
An evidence status of verified means its bytes still match the recorded or
human-acknowledged source version. It does not mean independent fact checking.
Outcomes without attached evidence are unverified, even when execution status
is completed. Changed or missing evidence produces review_required for linked
outcomes. A human acknowledgement keeps the original digest and records a reason.
Source change detection is based on content hashes. Question suggestions use explicit rules, not a trained value-of-information model. Feedback counts are human ratings, not an automatically generated quality score.
In Projetos, Analisar fontes uses the configured model to propose possible contradictions, temporal changes, complementary statements or uncertain relationships between project decisions and authorized source material. The service validates that cited quotations occur literally in their cited sources; that check does not establish that the model's interpretation is correct. Findings remain proposals until a person confirms or dismisses them with a reason. A report becomes stale when the relevant project or its sources change; review requires a fresh analysis. An unavailable source is shown explicitly and its derived details are withheld. Analysis of outcomes compares expected and observed results, but does not establish causality.
Analysis reads at most 100 sources and defaults to 65,536 source bytes, with a separate total prompt budget. Oversized inputs fail instead of being silently truncated. A native analysis has a three-minute timeout. A benchmark has a ten-minute overall timeout and a two-minute timeout per model request; cancellation also works while a local model is starting.
The Qualidade view runs six synthetic scenarios with the configured model,
using up to 12 model calls. It shows progress and supports cancellation. The same
configured model generates and judges the answers in separate calls; citation IDs and quoted evidence are also validated
against the synthetic sources and answers. A model judgment can be wrong or
inconclusive, so these results are not a universal quality score. The UI shows
up to 20 recent reports and compares runs only when suite version, fingerprint and
case IDs match. The reports are scoped to the active space and stored as JSON
under ops/answer-benchmarks, outside Markdown recall and Git memory ingestion.
They use the memory service's encryption policy: the normal encrypted vault
requires unlocked keys; explicitly enabled legacy plaintext mode follows its
existing policy. No private user memory is used as a benchmark scenario.
Native chat context, lexical retrieval, RAG and the standalone MCP reader share
tools/lib/memory-contract.mjs. Reads and searches default to current:
- Active, dated facts whose validity interval includes today are eligible.
- Superseded facts, archived facts, future/invalid dates and pending proposals do not enter current fact recall.
- Undated legacy material remains visibly labelled unverified.
- Other notes and logs remain source material or historical records. They are not promoted to canonical facts merely because retrieval found matching words.
mode: "history" explicitly includes archived records and proposals, preserving
their labels. mode: "asOf", asOf: "YYYY-MM-DD" selects dated facts valid on that
date and excludes proposals and undated material. Non-fact notes cannot establish
a dated fact interval. Search snippets retain provenance, classification, space
and version information; long fact chunks retain their individual status labels.
The index is refreshed before retrieval so a correction replaces old indexed
facts on the next query without waiting for a background rebuild.
MCP fact proposals use the client's default space, or the explicitly selected
fact document for sb_append. They remain pending until the person accepts them
in Hoje. Existing undated notes are not automatically rewritten as facts.
Records live at projects/<uuid>.md in personal memory or
spaces/<space>/projects/<uuid>.md in another space, with a Markdown wrapper and
a versioned JSON payload. The native memory service applies the existing
encryption and classification rules. MCP writes inherit the strongest
classification of the record and its readable sources. Sources must be within
the same project space; traversal, policy directories and symbolic-link sources
are rejected. A removed source does not prevent editing project metadata, and
its availability remains explicit.
A corrupt authorized record is reported as {id, error: "CORRUPT_PROJECT"} in
the list while healthy projects remain usable. Inaccessible records do not expose
their titles or contents. Repair corrupted files from a trusted previous version
instead of replacing them with an empty project.
Writes take an exclusive <uuid>.md.lock directory and check the revision while
holding it. This coordinates the app and MCP processes sharing the same local
vault. Independent Git clones do not share that filesystem lock; synchronize
and resolve repository conflicts before resuming concurrent work across clones.
This is not a distributed transaction protocol.
A process crash can leave a lock. The service fails closed with PROJECT_LOCKED
and never guesses that a lock is stale. Close all writers, back up the project,
verify that no process owns the operation, then remove only that project's empty
lock directory before reopening it. Do not delete the project to unlock it.
The private derived-learning registry keeps its existing owner-only access and retention behavior. Project outcomes do not automatically promote themselves into that registry or expire existing lessons.
Run npm run build, npm run typecheck, then npm test. Building first supplies
the packaged derived-learning module used by the existing MCP tests. New tests
cover shared native/MCP persistence, current/history selection, stale revisions,
human authority, changed and missing evidence, transitive artifact impacts,
encrypted classification inheritance, space isolation, corruption, context
budgets and path aliases.
Deterministic regression and renderer DOM interaction tests exercise the memory contract and user flows with test fixtures. They do not measure a real model’s answer quality. The Hoje indicator is named Retrieval checks to distinguish it from the synthetic answer benchmark. Visual inspection in an installed desktop application has not been completed.
An additional live Codex run, using only synthetic data, is recorded in the validation report. It verifies the provider path and retains the observed outputs; one run with a model judge does not establish reliability on real projects or a ranking between models.