Memory Layers¶

Akasha organizes agent knowledge into four cognitive layers, inspired by human memory systems.

The Four Layers¶

graph LR
    W[Working Memory<br/>TTL: seconds] --> E[Episodic Memory<br/>TTL: hours]
    E --> S[Semantic Memory<br/>Permanent]
    E --> P[Procedural Memory<br/>Permanent]

    style W fill:#fdcb6e,stroke:#f9ca24,color:#000
    style E fill:#00cec9,stroke:#81ecec,color:#fff
    style S fill:#6c5ce7,stroke:#a29bfe,color:#fff
    style P fill:#e84393,stroke:#fd79a8,color:#fff

Working Memory¶

Short-term scratch pad for active tasks. Auto-expires.

client.put("memory/working/task-42/context", {
    "current_step": "analyzing",
    "data_loaded": True,
}, ttl_seconds=60)

Path prefix: memory/working/
Default TTL: 60 seconds
Use case: Active task state, intermediate results, scratchpads

Episodic Memory¶

Event logs and interaction history. Medium retention.

client.put("memory/episodic/2024-01-15/user-login", {
    "user": "alice",
    "action": "login",
    "timestamp": "2024-01-15T10:30:00Z",
}, ttl_seconds=3600)

Path prefix: memory/episodic/
Default TTL: 1 hour
Use case: Conversation history, audit logs, event sequences

Semantic Memory¶

Permanent knowledge. Facts, learned patterns, distilled insights.

client.put("memory/semantic/customers/acme-corp", {
    "preferences": {"channel": "email", "tone": "formal"},
    "industry": "manufacturing",
    "tier": "enterprise",
})

Path prefix: memory/semantic/
TTL: None (permanent)
Use case: Knowledge base, learned preferences, domain facts

Procedural Memory¶

How-to knowledge. Workflows, procedures, learned sequences.

client.put("memory/procedural/deploy-canary", {
    "steps": ["build", "test", "deploy-10%", "monitor-5min", "promote"],
    "rollback_on": "error_rate > 1%",
})

Path prefix: memory/procedural/
TTL: None (permanent)
Use case: Standard operating procedures, learned workflows

Querying Across Layers¶

# All working memory
working = client.query("memory/working/**")

# All memories about a topic
about_acme = client.query("memory/*/acme-corp/**")

# Layer stats
stats = client.memory_layers()
# → {"working": 38, "episodic": 467, "semantic": 50, "procedural": 2}

Nidra: Memory Consolidation¶

The Nidra engine automatically consolidates memories:

1. Working memory entries that are read frequently → promoted to Episodic
2. Episodic entries with consistent patterns → promoted to Semantic
3. Expired entries → garbage collected

This mirrors how the human brain consolidates short-term memories into long-term storage during sleep.