Memory Consolidation

Memory Consolidation

Overview

Your agent accumulates many interactions, but most are redundant, overlapping, or partially inconsistent. Consolidation is the agent's "sleep phase": it compresses short-term experiences into long-term understanding. Ch4 frames this as four steps (Example 4-5):

1. Cluster related memories (cluster_by_topic) — group conversations

about the same project/incident by similarity.

1. Summarize each cluster (summarize_cluster) — replace "Monday: deadline

Friday; Tuesday: confirmed Friday; Wednesday: Friday again" with one fact: "Project deadline: Friday (confirmed 3 times)". Meaning preserved, redundancy gone.

1. Consolidate into permanent graph nodes (create_consolidated_memory).
2. Maintain provenance (maintain_provenance_chain) — keep the

DERIVED_FROM links so the agent can trace a belief back to the exact interactions that produced it.

Two disciplines from the chapter shape the implementation:

• Minimum cluster size (Example 4-13: "Need enough examples to

generalize"). A pattern derived from a single episode is not a pattern. Default minimum is 3.

• Sleep-time compute (Letta + UC Berkeley, Ch4): consolidation runs

during idle periods, not while a user waits. Shifting it off the response path cuts active inference cost ~5x and lets you pre-compute inferences (which tasks are at risk, given a deadline + dependencies) before anyone asks.

When to Use

• An agent with accumulating episodic memory that grows noisy over time
• DevOps incident memory: turn many similar 503-after-deploy incidents into

one durable Pattern node with a runbook

• Conversational assistants that repeat the same fact across sessions
• Any system that needs to answer "how do you know X?" with a provenance trace

Phrases: "consolidate memory", "summarize episodes", "sleep-time compute", "provenance chain", "compress experience into knowledge", "cluster incidents".

When NOT to Use

• One-shot / stateless agents — there is nothing to consolidate
• The synchronous response path — consolidation is a background/idle job
• Facts that must remain individually addressable (an audit log of distinct

events) — consolidation deliberately merges them

• Clusters that never reach the minimum size — keep the raw episodes; do not

fabricate a pattern from one example

Process

| Step | Input | Action | Output | Verification | |------|-------|--------|--------|--------------| | 1 | list of raw episodes | lib.cluster_by_topic(episodes, threshold) | list of clusters | related episodes group; unrelated ones split | | 2 | one cluster | lib.summarize_cluster(cluster) | one durable fact string | multi-episode summary carries a confirmation count | | 3 | all episodes | lib.consolidate(episodes, min_cluster_size) | list of ConsolidatedFact | clusters below min size are skipped | | 4 | a fact + episodes | lib.provenance_of(fact, episodes) | source episodes | round-trips; dangling links raise | | 5 | facts + idle worker | lib.precompute_inferences(facts, inference_fn) | {fact_id: inferences} | runs off the response path |

CLI: cluster, consolidate, scenario incident-consolidation, benchmark.

Rationalizations

| Agent rationalization | Documented rebuttal | |------------------------|--------------------| | "I'll consolidate on every turn so memory is always fresh." | Consolidation is expensive (clustering + summarization). The chapter's sleep-time-compute result is explicit: run it during idle periods, not the response path — ~5x cheaper active inference. On-turn consolidation pays the cost when the user is waiting. | | "One vivid episode is enough to make a pattern." | Example 4-13 sets `if len(cluster) = min-cluster-size times manufactures a "confirmed 3 times" consolidated fact. Clustering never executes episode text; the provenance chain is the audit path - inspect sources before trusting a consolidated node.

• Data exfiltration. Consolidation copies episode content into durable

nodes: sensitive details survive in summaries and DERIVED_FROM links long after raw episodes age out. Redact before consolidating; the skill makes no network calls and writes no files itself.

• Privilege escalation. No shell invocation, no eval. Consolidated facts

gain durable, trusted standing in the graph, and sleep-time jobs run unattended - bound the background job's write scope and gate what gets promoted from episode to knowledge.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly) by Anthony Alcaraz and Sam Julien — Ch4: Memory — "Consolidation: From Experience to Knowledge" (Example 4-5) and "Adding Memory to the DevOps Agent" (Example 4-13). Sleep-time compute: Letta + UC Berkeley research cited in Ch4. Narrative-fact extraction: HINDSIGHT (Latimer et al. 2025).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AnthonyAlcaraz
• Source: AnthonyAlcaraz/agentic-graph-rag-skills
• License: MIT
• Homepage: https://www.oreilly.com/library/view/agentic-graphrag/9798341623163/

Install and usage instructions live in the source repository linked above.