Route first: forgetting context is not truncation, quota, or file access
When OpeClaw forgets context or does not remember earlier turns, classify the symptom: a brand-new chat with no prior memory; the same chat drifting away from the opening; an answer that stops mid-generation; or a clear 401, 429, quota, or local-file error.
This page covers limited model context windows plus session management—how to structure prompts, split work, and retain critical constraints. Mid-output truncation belongs in the truncated-output guide; quota errors in the API key guide; workspace file reads in the local-folder guide.
Do not invent a fixed token ceiling from feel, and do not assume a named official “long-term memory” feature. Context capacity and session behavior follow the current model documentation and client-visible messages.
- New chat with no memory: treat it as a session boundary; restate constraints.
- Same chat forgets the opening: prioritize context eviction; shorten and pin rules.
- Stops mid-answer: switch to truncated / interrupted output.
- 401/429 or unread files: use the matching guide, not this one.
New chats have no memory: restate the boundary; do not assume inheritance
A new session is usually a new conversation boundary. If you need the previous role, forbidden actions, or project rules, write them at the start instead of assuming another window still “remembers.”
Keep must-follow constraints in a copyable short block: goal, prohibitions, output format, glossary. Paste that block first, then the current task. Whether a cross-session mechanism exists, and what it is called, depends on the current client.
Put reusable specs, samples, and folder notes in workspace files and cite paths in chat, instead of repeatedly pasting whole documents into history. For attach or unread-file issues, use the local-folder guide.
Forgetting the opening mid-chat: organize when context is squeezed out
Within one session the model works inside a limited context window. As the thread grows, earlier instructions, constraints, and details may leave the currently available context—so it looks like the opening was forgotten. That is not the same as a crash, and it does not justify guessing a fixed token number.
Do three things: resend still-valid critical constraints at the top; trim or archive unused long history; rewrite bulky material as “conclusion first, expand on demand” instead of dumping whole chapters. Capacity and truncation policy follow the current model documentation.
If the symptom is a mid-answer stop, half sentence, or unclosed code fence, that is an output-layer issue—use the truncated-output guide rather than only enlarging the prompt here.
- Critical-constraint checklist: repaste at each task start.
- Park long material in the workspace; cite only needed slices.
- One job at a time; finish before starting the next segment.
- If generation still stops mid-way, check the truncated-output guide.
Keeping critical information: prompt structure, task splits, and Skills limits
Retain critical information with structure, not by hoping the model “just remembers.” Recommended order: immutable rules → current goal → only this-turn material → output format. Shorter, earlier rules survive long threads better.
Split work: outline or acceptance criteria first, then execute section by section. Carry forward only confirmed conclusions into the next section so a limited window can relay summaries instead of reconstructing the entire history from scratch.
Skills fit repeatable workflows and tool capabilities; they do not replace confirming safety boundaries, factual limits, and output format each time. Install and enable belong in the Skills setup guide; enabled-but-failing calls (no permission, missing path, execution errors) belong in the tool/Skill call-failure guide. This page does not cover setup or runtime tool debugging.
Context truncation vs no session memory: routing table
Use the table for routing only. Exact window size, session control names, and any cross-session capability follow the current model documentation and client.
| Symptom | First classification | Next step |
|---|---|---|
| New chat remembers nothing from the previous one | Session boundary, not same-thread context | Repaste critical rules; keep materials in the workspace |
| Same chat forgets the opening over time | Earlier turns pushed out of the context window | Pin constraints, shorten history, split tasks |
| Answer stops halfway through generation | Output interruption / length or connection | Truncated & interrupted output guide |
| 401, 429, or quota message | Authentication or quota | API key and quota guide |
| Claims to read a file but content does not match | Workspace path or permission | Local folder / workspace guide |
Do not invent a fixed token ceiling or an official “long-term memory” feature name; follow the current model documentation and client-visible messages.
Minimum proof: confirm session issues before chasing other failures
Run a tiny rule check: in a new chat, write only “Every answer must start with the line OK,” then ask an unrelated question. If the first line is stably OK, the new session accepts current instructions. Expecting OK from a previous session without repasting usually fails—that matches a session boundary, not a broken install.
In a long thread, paste the same rule, then ask several unrelated follow-ups. If later turns drop the first-line rule, treat it as context eviction: pin the rule again and shorten history.
If quota, key, or proxy errors appear during the check, switch to those guides immediately. For install entry or release status, use the download page; general product questions belong in the FAQ.
Context & session management FAQ
Why does a new OpeClaw chat have no memory of the previous one?
In most cases a new session does not automatically inherit the full prior conversation. Restate critical constraints, role rules, and must-follow boundaries at the start of the new chat, or keep them in a workspace note you can cite. Whether any cross-session capability exists depends on the current client and model documentation—this site does not invent a fixed “long-term memory” feature name.
If OpeClaw forgets the opening later in the same chat, is the model broken?
The more common cause is a limited context window: earlier turns may fall outside the currently available context, which feels like forgetting the start. Shorten history, restate critical constraints at the top, and split the job. If the answer stops halfway through generation, use the truncated-output guide instead.
How do I keep critical information from getting lost?
Write immutable rules as a short checklist and paste it at the start of each task. Keep long source material in workspace files and cite only what you need, instead of dumping whole documents into the chat. Skills help with repeatable workflows; they do not replace restating your safety and format boundaries. Context capacity follows the current model documentation.
Should I use this guide for 401/429 or local file read failures?
Keep those paths separate. Clear 401/429 or quota messages belong in the API key and quota guide; path, permission, or unread workspace files belong in the local-folder guide; mid-answer truncation belongs in the truncated-output guide; enabled tools/Skills that still error belong in the tool-call-failure guide. This page only covers “still chatting, but it seems to forget earlier context.”
Use this with an OpeClaw workflow
Check the current OpeClaw download status first, then save this guide as part of your setup, review, or troubleshooting workflow.