Skip to main content

Memory Systems Overview

Aragora’s control plane uses layered memory to keep vetted decisionmaking grounded, persistent, and improvable across debates, workflows, and channels.

Core Components​

  • ContinuumMemory: Tiered memory with promotion/demotion rules and retention controls. (aragora/memory/continuum/core.py)
  • ConsensusMemory: Stores consensus outcomes, dissent, and similar-topic retrieval. (aragora/memory/consensus.py)
  • MemoryStream: Per-agent event timelines used for episodic recall. (aragora/memory/streams.py)
  • Knowledge Mound: Unified knowledge facade across multiple stores, with provenance and staleness tracking. (aragora/knowledge/mound/core.py)

Lifecycle at a Glance​

  1. Debate/workflow events are recorded into MemoryStream and ContinuumMemory.
  2. Consensus results and receipts are recorded into ConsensusMemory.
  3. The Knowledge Mound syncs across memory stores to provide unified query and governance controls.

Storage and Backends​

Memory defaults to local persistence for development. Production deployments should use durable backends (PostgreSQL/Redis/Weaviate) via Knowledge Mound configuration for multi-instance consistency.

ContinuumMemory supports hybrid search that fuses vector similarity and keyword matches using Reciprocal Rank Fusion (RRF). The keyword index is backed by SQLite FTS5 in the same memory database.

from aragora.memory import get_hybrid_memory_search

search = get_hybrid_memory_search(memory)
results = await search.search("circuit breaker pattern", limit=10)

Use KeywordIndex.rebuild_index() to (re)populate the FTS index after large ingestion batches or migrations.