Devotd Github
NewAI agents

tencentcloud/octop-memory v1.0.3 — Fix for long-session extraction failure and documentation corrections

tencentcloud/octop-memory 1.0.3 fixes the problem where candidate extraction in long sessions failed due to exceeding the context window by switching to batch processing and cursor persistence, and documents the Hermes installation path.

Repositorytencentcloud/octop-memory ↗

tencentcloud/octop-memory 1.0.3 fixes the problem where candidate extraction in long sessions failed due to exceeding the context window by switching to batch processing and cursor persistence, and documents the Hermes installation path.

GitHub description: Help your agent remember what matters. Keep preferences, carry context across sessions, and bring your memory to the next agent.

stars
★ 70
Forks
24
Language
Python
License
MIT
Latest release
v1.0.2
Read original source ↗View on GitHub ↗

tencentcloud/octop-memory released 1.0.3 on November 4, 2026. It was incorporated through the merge commit of PR #38 (dccc47c) in the GitHub repository, and the changes amounted to 5 files, 13 lines added and 7 lines deleted. This project is a memory system that carries over an agent's preferences and context even when the session changes. The package description advertises hierarchical recall, FTS search, and support for multiple backends, and the license is MIT.

The focus of this release, as recorded in the release notes (CHANGELOG.md), is the candidate extraction path that kept breaking in long sessions. Now, instead of putting all session events into a single request, it slices the events based on a single-batch event count (default 100) and a character budget (default 20,000) before calling the model. Previously such requests exceeded the context length and failed with 'length truncation/timeout' (#25). At the same time, a cursor meaning 'already extracted' is stored at extracted_ids/v1/<session> in {ns}_meta, so that after a restart processing resumes from meta. The incremental extraction path records the cursor with strict=True, and if the recording fails, the entire transaction is rolled back and the candidate record is cancelled as well. Events remain in a state awaiting retry rather than being silently skipped.

Auditing and page regeneration were also addressed. For past manual document audit records that lack an uncertain field, it no longer trusts the progress already compared, and returns active facts to the re-review target (#33). EntityPage regeneration compares page revision numbers so that it does not overwrite content the user edited during the model call. It also does not raise an exception even when a root or drive character is attached to an allowlist glob on Windows.

That said, 1.0.3 is a release centered on fixes and documentation corrections rather than feature additions. The Hermes installation guidance consolidates docs/integrations.md and the provider README into a single path, instructing users to install by specifying the full 40-character SHA with --ref from the plugins/hermes/octopmemory subdirectory within the repository, and to check the gateway interpreter with hermes --print-runtime-command. At the same time, it also states that this repository does not publish an independent plugin package on PyPI, so the existing commands return 404 (#10). This commit record contains no performance figures or user counts, so the extent of the improvement cannot be read as a number. It can at least be confirmed that the failure conditions and the installation procedure have been narrowed down in the documentation. Teams that dealt with long sessions in 1.0.2 have a reason to upgrade, and those that have not started using it yet can defer the decision.