Shared Persistent Memory

concept Apr 7, 2026

multi-agentmemoryknowledge-accumulationcoordination

Shared persistent memory is a coordination mechanism for multi-agent systems where agents communicate indirectly through a shared, durable knowledge repository rather than direct messaging.

Design in CORAL

CORAL implements shared persistent memory as a filesystem with three directories (source):

attempts/ — historical evaluations and solutions, keyed by commit hash (JSON records)
notes/ — observations, learnings, reflections as markdown files, organized into subdirectories by agent
skills/ — reusable procedures, tools, and implementation patterns with natural-language descriptions and executable artifacts

Each agent’s isolated workspace symlinks to this shared repository, maintaining a single source of truth. Agents read and write via CLI tools or direct filesystem access. Unique filenames per agent minimize write conflicts; no explicit locking is required.

Why filesystem

The filesystem design provides progressive disclosure — agents access only what they need without loading the entire knowledge base into context. It is also extensible: agents can create subdirectories to organize knowledge by category, and new artifact types can be added without schema changes. The symlink architecture preserves workspace isolation while enabling shared access.

Indirect coordination

Unlike peer-to-peer multi-agent systems (AutoGen, LangGraph, MetaGPT) where agents directly message each other, shared persistent memory enables indirect coordination. One agent’s discoveries influence another’s search through written artifacts: a technique documented as a skill, a failed approach recorded in notes, a high-scoring attempt visible on the leaderboard. This produces emergent behaviors — technique diffusion, spontaneous consensus, cross-referencing — without hardcoded communication protocols.

On Anthropic’s kernel engineering task, 36% of attempts use another agent’s commit as parent, and 66% of new records originate from cross-agent parents. On Polyominoes, 87% of rounds reference knowledge committed by other agents (source).

Connection to LLM wiki pattern

Shared persistent memory is a task-scoped application of the LLM wiki pattern. Both use persistent, structured knowledge stores maintained by LLMs that compound over time without forgetting. The LLM wiki pattern applies this to general knowledge management; CORAL applies it to evolutionary search, where the “wiki” is the agents’ collective understanding of what works and what doesn’t.

Backlinks

Agent Learning Loop CORAL's shared persistent memory extends this to multi-agent settings — multiple agents accumulate knowledge in a shared filesystem, and cross-agent parents inherit successful strategies. →

Autonomous Evolution

Shared persistent memory — agents need a durable knowledge store to accumulate insights across attempts
Heartbeat mechanism — periodic interventions prevent agents from drifting into local minima
Local verification — agents test candidates locally before consuming expensive evaluations
Knowledge accumulation — capturing reusable insights (not just progress logs) is causally linked to performance; disabling it degrades results by up to 18.6%

→

CORAL: Autonomous Multi-Agent Evolution Shared persistent memory as filesystem. Knowledge is structured into three directories — attempts/ (historical evaluations), notes/ (observations and reflections), and skills/ (reusable procedures). Each agent's workspace symlinks to this shared repository, maintaining a single source of truth while preserving workspace isolation. →

Cross-Agent Knowledge Transfer Cross-agent knowledge transfer is the mechanism by which one agent's discoveries influence another agent's search, without direct communication. In CORAL, this occurs through shared persistent memory — agents write artifacts that other agents later retrieve as context. →

Hermes Agent

Implements the LLM wiki pattern at the agent level — skills compound like wiki pages, memory persists like the raw layer
The 6 terminal backends directly implement the capsules concept — isolated environments that hibernate when idle
Skills system parallels CORAL's shared persistent memory — markdown files as accumulating knowledge
Context compression preserves structured summaries, similar to how the heartbeat mechanism in CORAL forces periodic reflection
Embodies the everything-is-text principle: all platforms, skills, memory, and config reduce to text

→

AI Agents

CORAL: Autonomous Multi-Agent Evolution — framework for autonomous multi-agent evolution on open-ended problems
Autonomous Evolution — paradigm shift from fixed evolutionary search to agent-controlled search
Open-Ended Discovery — problem class where objectives are clear but optimal solutions unknown
Shared Persistent Memory — filesystem-as-memory for multi-agent coordination
Heartbeat Mechanism — periodic interventions preventing local-optima drift
Cross-Agent Knowledge Transfer — indirect coordination through shared artifacts

→

Knowledge Management

Vannevar Bush Memex — the 1945 vision of a personal knowledge store with associative trails
Shared Persistent Memory — filesystem-as-memory for multi-agent coordination (CORAL)

→

Open-Ended Discovery

Fixed evolutionary search: FunSearch (Romera-Paredes et al., 2024), AlphaEvolve (Novikov et al., 2025), OpenEvolve, ShinkaEvolve, EvoX — LLM proposes mutations, external algorithm manages selection and population
Autonomous evolution: CORAL (Qu et al., 2026) — agents control the full search process, accumulating knowledge through shared persistent memory

→

Session as Context Object

Managed Agents Architecture: the system that implements this pattern
Context window compression: one of the transformations the harness can apply to events fetched from the session
Shared persistent memory (CORAL): a related pattern for multi-agent systems — agents coordinate through a durable filesystem rather than direct messaging. The session log serves a similar role for a single agent across time.

→