Receive
Give new claims a holding area.
Incoming knowledge stays in a separate review state, with its source and arrival time attached.
Learning isolation
New information should earn its place.
Give new information a review path before it becomes trusted context. Learning Isolation in Akasha keeps incoming claims separate while configured checks assess their consistency, sources, and confidence.
Learning isolation
New information.
Earned trust.
Give a new claim somewhere to be checked before it becomes accepted knowledge.
Knowledge review
Incoming claimLaunch date · September 20 Awaiting review
Supported updateLaunch date · September 24 Source · Signed release plan
Your accepted knowledge stays separate.
Two claims arrive. Neither is trusted yet.
Incoming knowledge enters a separate review state. Your accepted records remain unchanged.
From claim to trusted context
Check the claim before changing the record.
An imported report says a customer changed their requirements. Existing records say something else. Learning Isolation holds the new claim for checks and review, instead of immediately replacing what the agent already knows.
Check
Look for reasons to pause.
Validation checks look for inconsistencies, unusual inputs, and signs of poisoned information before a claim moves forward.
Consolidate
Find what is worth keeping.
The consolidation pipeline groups related experiences and identifies candidate facts for longer-term memory.
Verify
Check the fact in context.
Rules check candidate facts for contradictions, time relationships, and valid structure before required verification is satisfied.
Accept
Control what joins trusted memory.
Acceptance follows validation and trust requirements. Items can be approved, rejected, or returned for another round of checks.
Recover
Keep a route back.
Use saved checkpoints to restore selected model or embedding components if a later update needs to be reversed.
Developer integration
Set the standard for new knowledge.
Use the Rust quarantine library to configure how incoming knowledge is reviewed. Choose capacity and test requirements, then integrate those rules with the sources and checks your application uses.
Configure knowledge review in the Rust enginerust
use akasha_learning_quarantine::config::QuarantineConfig;
let mut review = QuarantineConfig::default();
review.require_all_tests = true;
review.max_capacity = 5_000;
review.validate()?;
// Supply this configuration when integrating the quarantine engine.
// Your application supplies the knowledge and verification context.Under the hood
Clear checks. Controlled acceptance.
Configure how incoming claims are tested and when they can be accepted. The quarantine library exposes review rules and trust settings for developers integrating a knowledge pipeline.
Quarantine & consolidation
- Knowledge lifecycle
- Receive → validate → assess trust → accept or reject
- Validation checks
- Consistency · cross-verification · anomalies · poisoning signals
- Domains
- Factual · Procedural · Semantic · Episodic · External
- Learning gate threshold
- 0.5 (default)
- Minimum confidence
- 0.6 (default)
- Verification required
- Enabled by default in consolidation
- Experiences per batch
- Up to 1,000 by default
- Isolated snapshots
- MVCC — versioned access during review
Rust integration
- Library
- akasha-learning-quarantine
- Configuration
- QuarantineConfig
- Review capacity
- max_capacity
- Test policy
- require_all_tests · min_tests_passed
- Presets
- default · strict · permissive
- Integration
- Application-supplied knowledge and verification context
Be deliberate about what your agent learns.
Build a review path for new information before it joins the knowledge your agent relies on.