Ananke.Learning 0.8.5

Ananke.Learning

👁 NuGet
👁 License

Empirical memory and experience-based learning for .NET AI agents — commit observations, recall patterns and heuristics, reinforce or contradict beliefs, run offline learning cycles, track episodes, and package learned behavior for transfer.

Part of the Ananke framework.

Install

dotnet add package Ananke.Learning

This package depends on Ananke.Orchestration and Ananke.Abstractions.

Key concepts

Observation ----> CommitAsync ----> EmpiricalEntry (stored in IEmpiricalMemory)
 |
Query -----------> RecallAsync ----------| vector similarity + tag overlap
 |
Outcome ---------> ReinforceAsync -------| prediction-error -> confidence/strength
 |
Background ------> LearnAsync -----------+ decay, curiosity, simulation, consolidation

An empirical entry represents a learned pattern, skill, or heuristic. Entries carry structured semantic descriptions, tags, confidence, observation history, affective signals, optional episode linkage, and optional entity scope. They are committed during agent execution, recalled by similarity, reinforced or contradicted by outcomes, and periodically swept by the offline learner.

What is included

Core interfaces

Built-in implementations

Data types

Agent tool integration

Quick start

Commit and recall

using Ananke.Learning.EmpiricalMemory;
using Ananke.Orchestration.Knowledge.Embeddings;

var embedder = new InMemoryEmbedder();
var memory = new InMemoryEmpiricalMemory(embedder);

// Commit an observation
var entry = await memory.CommitAsync(new EmpiricalEntry
{
 Id = Guid.NewGuid().ToString("N"),
 Kind = EmpiricalKind.Pattern,
 Tags = ["strategy", "center-control"],
 Source = "manual-observation",
 Description = new SemanticDescription
 {
 Summary = "Placing pieces in the center column creates more threats",
 SemanticTags = new Dictionary<string, float>
 {
 ["strategy:center-control"] = 1.0f,
 ["domain:connect4"] = 0.8f
 }
 },
 Confidence = 0.5f,
 ObservationCount = 1,
 Evidence = [],
 FirstObserved = DateTimeOffset.UtcNow,
 LastObserved = DateTimeOffset.UtcNow
});

// Recall by similarity
var matches = await memory.RecallAsync(
 "what strategy works for opening moves?",
 new RecallOptions { TopK = 5 });

Reinforce with outcome

await memory.ReinforceAsync(entry.Id, new Reinforcement
{
 Reward = 1.0f,
 NewEvidence = ["won the game after using this strategy"],
 Source = "game-outcome"
});

Run offline learning

using Ananke.Abstractions.Agents;
using Ananke.Learning.Offline;
using Ananke.Orchestration.Knowledge;

var knowledgeStore = new InMemoryKnowledgeStore(embedder);
var learner = new OfflineLearner(
 memory,
 embedder,
 knowledgeStore: knowledgeStore,
 options: new OfflineLearnerOptions
 {
 ExplorationBatchSize = 5,
 ConsolidationMinObservations = 10,
 ConsolidationMinStrength = 0.8f,
 });

var result = await learner.LearnAsync();

// result.Decayed, result.Explored, result.Consolidated

Entity-scoped memory

using Ananke.Learning.EntityMemory;

var provider = new EntityMemoryProvider(
 conversationMemory,
 memory,
 episodeStore,
 knowledgeStore);

var customerMemory = provider.GetOrCreate("customer-123");
var customerPatterns = await customerMemory.Empirical.RecallAsync("billing dispute");

Skill export/import

using Ananke.Learning.Skills;

var packager = new SkillPackager();
await using var writer = new JsonSkillPackageFormat().CreateWriter(stream);

var export = await packager.ExportAsync(
 new SkillExportOptions { Name = "support-playbook", Domain = "customer-support" },
 memory,
 writer,
 episodes: episodeStore);

Expose as agent tools

var toolkit = EmpiricalMemoryTools.Create(memory);

// Register toolkit.Tools with an AgentJob so the LLM can
// recall patterns, commit new observations, and reinforce entries
// through natural-language tool calls.

Learning lifecycle

1. Commit — Agent observes something and commits it as an EmpiricalEntry with semantic tags and an embedding.
2. Recall — On future decisions, the agent recalls similar entries by vector cosine similarity and tag overlap.
3. Reinforce — When outcomes arrive, entries are reinforced (positive or negative). Prediction-error mechanics update confidence, strength, variance, and affective signals.
4. Contradict — Entries disproven by evidence have their confidence reduced.
5. Episode tracking — Completed action sequences can be persisted and replayed for reward propagation and packaging.
6. Offline sweep — A background process runs periodically:
   • Decay removes entries whose strength has fallen below threshold.
   • Exploration selects curious (high-variance, low-observation) entries and tests them via ISimulationSource.
   • Consolidation promotes mature, high-confidence entries into IKnowledgeStore for long-term retention.
7. Packaging — Stable skills and supporting episodes can be exported and imported between agents or environments.

Package boundaries

• Ananke.Learning owns empirical memory, episodes, offline learning, feature tracking, entity-scoped facades, and skill packaging.
• Ananke.Orchestration.Knowledge supplies the long-term knowledge store used during consolidation.
• Ananke.Orchestration supplies ToolKit integration for exposing learning capabilities to agents.

Related packages

Worked example

See the Connect4 learning demo in the repository for an agent that learns strategy through self-play, empirical memory, and offline consolidation.

Documentation

Full docs, demos, and architecture: github.com/sevensamurai/Ananke

License

Apache 2.0