Ananke.Learning 0.4.0

Ananke.Learning

Empirical memory and skill learning for .NET AI agents — commit observations, recall by semantic similarity, reinforce with prediction-error feedback, and run offline learning sweeps that decay weak beliefs, explore curiosity-driven hypotheses, and consolidate mature patterns into long-term knowledge.

Part of the Ananke framework.

Install

dotnet add package Ananke.Learning

This package depends on Ananke.Orchestration (which transitively includes 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 single learned observation — a pattern, skill, or heuristic — with semantic tags, an embedding vector, confidence, strength, and affective signals (valence/intensity). Entries are committed during agent execution, recalled by similarity, reinforced when outcomes arrive, and periodically swept by the offline learner.

What is included

Core interfaces

In-memory implementations

Data types

Agent tool integration

Quick start

Commit and recall

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

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

// Commit an observation
var entry = await memory.CommitAsync(new EmpiricalEntry
{
    Kind = EmpiricalKind.Pattern,
    Description = "Center control leads to more wins",
    Semantic = new SemanticDescription(
        Tags: [new("strategy", "center-control", 1.0f)],
        Summary: "Placing pieces in the center column creates more threats",
        EmbeddingSourceText: "center column control strategy"),
});

// 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,
    Evidence = "won the game after using this strategy",
});

Run offline learning

using Ananke.Orchestration.Knowledge;

var knowledgeStore = new InMemoryKnowledgeStore(embedder);
var learner = new InMemoryOfflineLearner();

var result = await learner.LearnAsync(
    memory,
    knowledgeStore,
    new OfflineLearnerOptions
    {
        DecayRate = 0.02f,
        ExplorationBatchSize = 5,
        ConsolidationMinObservations = 10,
        ConsolidationMinStrength = 0.7f,
    });

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

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. 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.

Related packages

Worked example

See the Connect4Demo — an agent that learns Connect Four strategy through self-play, empirical memory, and offline consolidation.

Documentation

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

License

Apache 2.0