Skip to content
archive

ai-agent-memory-persistence

wiki · · 1 min read

AI Agent Memory and Persistence Patterns

A recurring theme across multiple sources is the challenge of giving AI agents persistent, structured memory. Stateless chat sessions are fundamentally limited — agents need long-term context that survives across interactions and compounds over time.

The Problem

AI coding agents (Claude Code, Codex, Copilot) and personal agents (Hermes) operate in discrete sessions. Each session starts nearly fresh — the agent has access to the immediate prompt and file context, but no accumulated knowledge from previous sessions beyond what's explicitly provided via system prompts or config files.

Emerging Solutions

File-Based Memory (agentmemory): rohitg00/agentmemory demonstrates persistent memory storage for AI coding agents using a structured file system. Agents read and write to designated memory files between sessions, creating a form of long-term contextual continuity.

Vault-Based Intelligence (Hermes): The FACorreia vault's raw → wiki → concepts pipeline is itself an agent memory system. Every ingested URL, tweet, or note becomes structured knowledge that future agent sessions can reference. The vault IS the agent's long-term memory.

MCP-Connected Vaults (Obsidian + Claude Code): The self-writing architecture connects agents directly to vaults via MCP, eliminating the file-copy-paste barrier. The agent operates on the vault natively, treating it as its own persistent workspace.

Tamper-Proof Audit Chains (psy-core): For production systems, memory integrity matters. psy-core provides hash-chained, HMAC-signed audit trails for agent operations — ensuring the agent's memory cannot be silently modified or corrupted.

The Convergence Point

The pattern emerging across these approaches is clear: agent memory should be (1) file-based, (2) human-readable Markdown, (3) organized in a navigable structure, and (4) accessible through standard protocols (filesystem, MCP). The vault-as-memory pattern that Hermes implements is a concrete step toward this.

Connected Concepts

  • Self-Writing Knowledge Architecture
  • Knowledge Base
  • Hermes Routing Policy
  • Swift Concurrency Evolution 2026

Sources

  • GitHub - rohitg00/agentmemory: #1 Persistent memory for AI coding agents based on real-world benchmarks · GitHub
  • <a href="/bookmarks/cyrilxbt-on-x:-"your-obsidian-vault-can-now-write-back-to-itself.-here's-the-architecture-nobody's-talking-about"-/-x" class="text-primary underline underline-offset-2">CyrilXBT on X: "Your Obsidian Vault Can Now Write Back to Itself. Here's the Architecture Nobody's Talking About" / X
  • HermiesVault MVP Specification
  • Hermes Routing Policy