One-Line Summary: Cross-session memory strategies decide what an agent remembers between conversations — the durable artifacts (configuration files, summaries, trajectories, adapters) and the policies for writing, retrieving, and aging them; this is one of the highest-leverage UX dimensions of any harness.

Prerequisites: Harness-owned memory, AgentDB and vector stores in harnesses, ReasoningBank

What Is Cross-Session Memory?

A session is bounded: it begins, you talk, it ends. Cross-session memory is everything that survives that boundary. The categories:

  1. Configuration / declared memory (CLAUDE.md, .cursorrules): the user writes these; the harness reads them on every session.
  2. Conversation transcripts: the harness keeps the raw transcript; later sessions can resume or grep.
  3. Compacted summaries: the harness summarizes long sessions into short notes to fit future contexts.
  4. Vector-indexed memories: extracted facts, decisions, error→fix pairs.
  5. Trajectory store (ReasoningBank): full rollouts indexed by signature.
  6. Parametric adapters (micro-LoRA): patterns baked into model weights.

A mature harness uses all of them, with explicit policies about what flows where.

How It Works

Consider a typical day's pipeline:

  • Session ends: Trigger compaction. The harness writes a 1–3-paragraph summary plus extracted facts to the per-project memory.
  • Outcome labeled: User signal or test result tags the trajectory.
  • Trajectory writes to ReasoningBank: With outcome label.
  • Periodically: SONA pattern extractor runs; new patterns emerge from clustered trajectories.
  • Periodically: LoRA adapters are retrained from cumulative trajectories.

Next session:

  • Read configuration: CLAUDE.md files concatenated into system prompt.
  • Resume option: User chooses to continue or start fresh.
  • Per-turn retrieval: Vector store and ReasoningBank are queried; relevant memories injected.
  • Adapter loaded: If available for this project.

Why It Matters

Cross-session memory is what makes an agent feel like a coworker rather than a tool. Coworkers remember what you talked about last week; tools don't. A harness with strong cross-session memory carries continuity — patterns the user didn't have to re-explain, decisions that didn't have to be re-justified, errors that don't have to be re-debugged.

The cost: cross-session memory is the hardest engineering surface in a harness. What to remember, what to forget, when to retrieve, how to summarize, how to age — every choice has a UX cost if wrong.

Key Technical Details

  • Forgetting is essential: Memory that grows without bound becomes noise. Aging policies (delete after N days unless reinforced; summarize older entries) are part of the design.
  • Per-project scoping is the safe default: Cross-project leakage of memories is a privacy and confusion liability.
  • Resume-vs-new decision: Users want both. A resume that replays the wrong session is worse than starting fresh; starting fresh on a 20-message conversation is also bad. Harness should default smartly and let user override.
  • Compaction quality matters: A bad summary is worse than no summary because it confidently misremembers. Test summaries against original trajectories periodically.
  • Conflict resolution: Memory says X; current observation says not-X. The agent has to decide which to trust. Usually current observation wins, but flagging conflicts to the user is the better UX.
  • User-visible memory controls: A user should be able to inspect, edit, and delete memories. Without this, memory is opaque and untrustable.

How Harnesses & Frameworks Implement This

Harness / FrameworkCross-session memory
Claude CodeCLAUDE.md + transcript persistence + --continue/--resume
Claude Agent SDKAll layers programmatically
rufloAll six categories first-class
LangGraphCheckpointers + MemoryStore pattern
AutoGenAdd-ons (mem0, memori) for the latter categories
CrewAIBuilt-in short-term + long-term memory
OpenAI Agents SDKConversation state + DIY for rest
Codex CLIAGENTS.md + transcript
Cursor.cursorrules + chat history + codebase index

Connections to Other Concepts

  • harness-owned-memory.md — The category.
  • agentdb-and-vector-stores-in-harnesses.md, reasoning-bank.md, sona-self-learning-neural-patterns.md, micro-lora-adapters-at-the-harness-layer.md — The memory layers.
  • memory-portability-across-harnesses.md — Cross-harness rather than cross-session.
  • ../../ai-agent-concepts/03-memory-systems/long-term-memory.md — Foundational coverage.

Further Reading

  • ruvnet, ruflo USERGUIDE — Memory pipeline overview.
  • LangGraph, Memory Patterns — Practical guide to memory in LangGraph.