v1.6.10 Reliability & Observability Improvements2026-11-18
v1.6.10 Reliability & Observability Improvements
Release Date: 2026-11-18
Changes: v1.6.9 → v1.6.10
Pull Requests: #6744
---
Summary
This release focuses on making Cognee more reliable and easier to debug: Postgres connection pools are tightened to reduce runaway connections and long waits, LLM call capture now records nested follow-ups and caller information for better observability, and the document chunking path is hardened to avoid incorrect chunk reuse. CI/test sharding was also expanded to reduce flakiness.
Highlights
- Safer Postgres per-dataset connection pools: lower default overflow and bounded wait times to avoid huge connection spikes and long stalls.
- Improved LLM call capture: records follow-up (nested) LLM calls and includes caller metadata for clearer debugging and usage analysis.
- Chunking robustness: stricter validation and centralized stamping of chunks so stored chunks are reliably identified and reused when possible.
- CI split into more shards to lower flakiness and speed up test feedback during development.
- Documentation updated for a retired Anthropic model name (now recommends claude-sonnet-5-5).
Breaking Changes
- Postgres per-dataset pool behavior: Defaults for per-dataset Postgres connection pools have been tightened (smaller max overflow and explicit pool wait behavior). If your deployment relies on very large, short-lived spikes of concurrent writes to per-dataset engines, you may see more immediate 'pool exhausted' errors. Action required: review and, if needed, increase your VECTORPOOLARGS or dataset-specific pool settings and/or raise pool_timeout to suit your workload.
New Features
- LLM call capture follow-ups: The LLM call capture feature (which writes one JSONL record per LLM call Cognee makes) now also captures follow-up / nested LLM calls and records who called them (a "caller" field). What it does: you get a clearer, traceable record of both the original LLM request and any internal subsequent calls. Why it matters: easier debugging, cost analysis and observability of complex multi-step LLM flows.
Improvements
- Per-dataset Postgres pool defaults tightened: The default ceiling for per-dataset Postgres pools (used by the pgvector backend and other per-dataset DB engines) has been reduced so each dataset uses fewer excess connections. The configuration comments now also call out a pool_timeout value (how long an operation waits for a free connection before failing). What it does: reduces the chance of opening too many DB connections and of long, unexpected stalls. Why it matters: more predictable deployments and lower database resource pressure in multi-tenant setups.
- Stricter chunking checks: The document chunking logic (which splits documents into smaller 'chunks' for indexing and retrieval) now performs tighter validation of chunks. What it does: catches invalid or mismatched chunks earlier and avoids silently accepting chunks that could corrupt update behavior. Why it matters: fewer mysterious rebuilds or data inconsistency when updating documents.
- Centralized chunk stamping: Chunk objects are now stamped centrally with the chunker identifier (the marker that shows which chunking rule produced the chunk). What it does: ensures stored chunks reliably carry the origin id instead of relying on distributed stamping. Why it matters: improves correctness when the system decides whether existing chunks can be reused during incremental updates.
- Docs: Anthropic model name updated: references to the retired model name 'claude-3-5-sonnet-20241022' were replaced with 'claude-sonnet-5-5' in documentation. What it does: prevents confusion from referencing removed models. Why it matters: keeps setup instructions accurate and reduces wasted time when configuring Anthropic providers.
Performance
- Lower connection overhead for per-dataset Postgres engines: By reducing the pool overflow ceiling and clarifying pool timeouts, typical deployments will open fewer concurrent DB connections. Practical impact: reduced DB resource usage and lower contention on hosts that serve many datasets; extreme spikes will fail faster instead of blocking for long periods.
Bug Fixes
- Fixes to unit tests and CI stability: Several failing unit tests were fixed and the test suite was reorganized to improve reliability (see CI changes below).
- Chunking-related bug fixes: Tightened chunker checks and centralized stamping resolve cases where chunks without proper identifiers could cause update failures or inconsistent state.
- Postgres pool and wait behavior fixes: Addressed issues where per-dataset pools could grow too large or block callers for long periods without failing fast.
Technical Changes
- CI/test sharding increased: Integration tests were repartitioned (Retrieval -> 5 shards, Documents & Tasks -> 6 shards, Infrastructure -> 4 shards) and unit tests split into 4 shards in cross-OS CI. What it does: reduces test flakiness and evens out execution time per shard. Why it matters: faster, more reliable CI feedback for developers.
- New/updated observability internals: new module files added to capture and write richer LLM call records (including follow-ups) and associated tests added to cover this functionality.
- Internal: MCP version bumped and multiple test files adjusted/added to reflect the new behavior and to lock down regressions.
Compatibility
| Component | Supported / Required |
| --- | --- |
| Python | >=3.10,<3.15 |
| pydantic | >=2.13 |
| litellm | >=1.102.2,<2.0.0 |
| fastapi | >=0.116.2,<1.0.0 |
| sqlalchemy | >=2.0.39,<3.0.0 |
| lancedb | >=0.39.0,<0.40.0 |
| ladybug | ==0.22.3 |
---
— The Cognee Team · 2026-11-18
v1.6.9 — New Connectors & Integrations2026-11-18
v1.6.9 — New Connectors & Integrations
Release Date: 2026-11-18
Changes: v1.6.8 → v1.6.9
Pull Requests: #6639, #6644, #6739, #6742
---
Summary
This release focuses on expanding integrations and stabilizing ingestion and embedding behavior. We added multiple new connectors (Slack rework, Dropbox, Vercel, Telegram, Evernote, Asana improvements) and a hermes-agent integration, plus several reliability and embedding fixes so data ingestion and retrieval are more dependable.
Highlights
- Major connector additions and Slack host refactor for improved Slack integrations.
- Hermes-agent integration: a packaged agent for capture, recall, and runtime workflows.
- Embedding and storage reliability fixes, including rejecting short embedding responses and bounded pending batches.
- Simplified and consistent document chunking (the logic that splits content for indexing) to avoid inconsistent splits.
Breaking Changes
- Slack history import removed: The older Slack history import pathway was dropped and replaced by the new Slack connector and host modules. What this means: if you depended on the old Slack history import API, you must migrate to the new Slack connector (connectors/slack) and update any custom wiring to the new host/adapter interfaces.
- Connectors packaging and ports: Several connectors were added and some Slack internals were moved into new host modules. What this means: if you vendor or patch connector internals, review the connectors/* layout and update imports to the new locations.
New Features
- Slack host rework and new Slack source: The Slack connector was ported into new host modules and a new direct Slack source was added, replacing the old Slack history import. What it does: provides a modernized Slack integration path (message sync, member linking, webhooks) maintained inside the connectors package. Why it matters: Slack ingestion and interactions are more maintainable and testable; if you previously used the removed Slack history import, migrate to the new Slack connector (see connectors/slack).
- Dropbox connector: A new Dropbox connector (client, connector, selection UI and source) was added. What it does: lets you ingest files from Dropbox into Cognee for indexing and retrieval. Why it matters: expands supported sources so you can pull documents from Dropbox for memory and search.
- Vercel connector: Added a Vercel connector and wiring for ingesting project data. What it does: provides a source implementation to capture Vercel-related artifacts. Why it matters: you can now include Vercel project data in your Cognee memories and searches.
- Telegram connector: Added a Telegram connector. What it does: captures Telegram message streams and integrates them into Cognee. Why it matters: expands chat platform coverage so Telegram conversations can be searched and recalled.
- Evernote connector: Added an Evernote connector. What it does: imports notes and notebooks from Evernote. Why it matters: users with Evernote content can include those notes in Cognee memories for retrieval.
- Hermes-agent integration: A first-class hermes-agent integration was added (provider, runtime, backend, capture policy, and tools). What it does: supplies an opinionated agent for capture, buffering, recall notices and runtime orchestration. Why it matters: simplifies running automated capture-and-recall workflows and local agent-based integrations with Cognee.
Improvements
- Default chunker unified and resolved once: The logic that decides how to split documents for indexing (the "chunker") is now resolved in one place and used consistently. What it does: prevents inconsistent splitting across code paths. Why it matters: search and embedding behavior becomes more predictable and you get more consistent retrieval results.
- Observability: LLM call capture: The system now records LLM (large language model) calls for observability. What it does: logs and telemetry include LLM calls invoked during operations. Why it matters: helps you diagnose when queries trigger LLM calls and measure usage.
- Accept graph model as JSON schema: The cognify pathway now accepts graph model definitions provided as JSON schema. What it does: makes it easier to supply a schema for graph-style data extraction. Why it matters: simplifies integration with systems that produce JSON schema graph definitions.
- Docs and docstrings strengthened: Many public APIs had their documentation and docstrings clarified (retrieval behavior, remember/add doc details, error cases). What it does: clarifies when retrievers are bypassed and permission-check behavior. Why it matters: easier to understand expected behavior and reduce integration bugs.
- Testing and CI: Test suites were split and pinned, and many connector unit/integration tests were added and stabilized. What it does: improves continuous integration reliability for the repository. Why it matters: reduces false CI failures and speeds safe releases.
Performance
- More reliable ingestion restarts for Asana: listings that previously stalled when offsets expired now restart correctly, reducing missed items during ingestion.
- Bound pending batches in storage: pending batch sizes are now bounded to prevent unbounded memory growth while buffering writes.
- Reduced redundant chunker resolution: resolving the default chunker only once cuts overhead and avoids repeated decisions during ingestion.
Security
- No new security advisories in this release. Note: Slack integration code was reorganized and signature verification utilities were adjusted as part of the Slack host refactor; if you rely on custom Slack hooks, review the new connector implementation.
Bug Fixes
- Reject short embedding responses for pgvector storage: very short or malformed embedding responses are rejected instead of being stored, preventing corrupted vectors and index errors.
- Recheck skipped rows in pgvector flows: rows previously skipped during a check will be re-evaluated to avoid losing items from the embedding index.
- Count Telegram failures once: failure metrics for Telegram ingestion were deduplicated so failure counts are accurate.
- Restart Asana listings when offset expires: Asana ingestion now restarts a listing job whose stored offset expired, avoiding stuck ingestion.
- Drop hardcoded connection limit: removed a hardcoded connection cap that caused connection handling problems in some environments.
- Scope embedding request splitting: refined how embedding requests are split so the right requests are sent together, reducing incorrect or partial embedding uploads.
- Tell apart errors that share the first line: error reporting was improved so exceptions with the same first-line message are distinguished correctly.
- Fix failing CI import errors and test robustness changes: multiple test fixes to ensure new tests run correctly in CI and avoid platform-specific failures (e.g., pytest-timeout SIGALRM usage).
Technical Changes
- Large refactor and porting of Slack host modules and adapters to the connectors package (new module layout and router/sync/adapter implementations).
- Major new connector code trees added: Dropbox, Vercel, Telegram, Evernote, Asana test/connector updates and many unit/integration tests for those connectors.
- Hermes-agent: extensive new code (provider, backend, runtimeenv, httpbackend, capture policy, tools) to support agent-based capture and recall workflows.
- Test and CI reorganization: tests split into groups, new GitHub workflow files added/updated to improve integration and e2e test coverage.
- Code quality: import sorting, docstring strengthening, and several refactors for clearer error handling and typing improvements (e.g., Asana auth and restart typing).
Compatibility
| Component | Supported / Required |
| --- | --- |
| Python | >=3.10,<3.15 |
| pydantic | >=2.13 |
| litellm | >=1.102.2,<2.0.0 |
| fastapi | >=0.116.2,<1.0.0 |
| sqlalchemy | >=2.0.39,<3.0.0 |
| l
v1.6.8 Integrations & Connectors2026-11-12
v1.6.8:集成与连接器更新
这版重心在两块:一是把集成做厚做稳,二是给后台任务补上可靠的基础设施。新增了完整的 Slack 连接器和 @cognee 应用,团队可以直接在 Slack 里向 Cognee 提问、存资料,不必切换工具。同时引入基于 RabbitMQ 的向量嵌入工作线程,把嵌入生成从主服务中拆出来,大批量或突发的导入不再卡住主流程,任务失败也能重试和分发。
Google 系连接器继续修补:Drive 现在能读取旧版 .xls、.ppt 等遗留格式,并回头重新处理此前同步失败的文件;Gmail 与日历改进了 API 调用与重试逻辑,减少需要手动重跑的情况。POST /add 新增 dlt_config 参数,可在导入时传入增量同步配置;Agent 检索器的工具输出字符上限也改为构造参数,便于按场景调节。
可靠性方面,更新与删除流程改为并发分批清理,存储文件按位置加锁,多任务碰不同后端时互不阻塞;审计账本的写入也做了修正,避免历史记录丢失或乱序。
v1.6.72026-11-06
v1.6.7(2026-11-06):入库更严,检索更快。
主要破坏性变更在数据入库侧:原始数据库连接串和经 HTTP 的服务器文件路径现在默认被拒绝,依赖这类连接串的旧流程会直接失败;ACCEPTLOCALFILE_PATH 也开始把关 s3:// 输入,原先依赖宽松行为的用户需按文档显式开启,否则相关入库会被挡下。拒收提示也统一了,能看出到底触发了哪条规则。
检索方面,带类型或命名空间限定的查询在分块搜索、引号查询里都能稳定生效,不再无谓走 LLM 路由;LLM 路由与范围路由的判定逻辑拆成了独立规则,行为更可预期。"手气不错"式推荐会跳过已禁用的数据类型。
性能上,PGVector、Turso、LanceDB 的查询打分改为批量并行,提速明显。
其余改动:文件入库被拒时返回 403 而非 415;新增 DLT 列图锚点字段,账本与图的对账更可靠;修复 LanceDB 压缩迁移、图 upsert、会话/遗忘操作及 SDK 导出头保留等问题。
v1.6.6 — Conflict resolution & stability2026-10-29
v1.6.6 主打冲突处理的可靠性。新增持久化的事实冲突解决 API,可在接口层记录并应用裁决结果,跨会话保持一致;冲突审查改为默认开启,相关修复确保用户在冲突上的决策(choices)在编辑、删除、重建等操作中被正确退役,不再留下过期判断。向量方面支持直接在不同向量库之间复制已有 embedding,迁移或复制时无需重新调用模型,省时省成本。稳定性上,Redis 列表与上下文删除改为原子或原地操作,权限与 KMS 报错不再被吞掉;无效的 DLT 键会直接报错而非静默跳过,编辑时相关状态失效失败也会中止,避免隐性数据不一致。另外,DLT 列统一按值节点处理、日期单元格只表示一次,修正了读取数据行时的重复与精度问题。
v1.6.6rc1 — History controls, connectors, and sync hardening2026-10-23
v1.6.6rc1 更新聚焦历史记录控制、新连接器与同步稳定性。forget 命令与 API 新增 clean_history 选项,可单独清除搜索历史;服务端也开放了用户自助删除搜索记录的接口,便于隐私清理。新增 Obsidian、Asana、OneNote 三个连接器,Drive 与 Asana 同步更快且支持断点续传,图与索引操作也减少了冗余开销。
需留意三处破坏性变更:删除数据集会连带清除关联的搜索与对话历史,依赖历史做备份的用户请先导出;run_info 不再内嵌完整 data,相关代码需改为按 data id 取原始数据(本版附带迁移);清除历史失败的 delete/forget 操作会直接报错,调用方需做好权限与异常处理。另外预算耗尽与付费类错误统一抛出 LLMPaymentRequiredError,涉及计费的客户端应捕获处理。
v1.6.5 — Faster trace processing & richer code-graph docs2026-10-16
v1.6.5 主打追踪处理提速与代码图文档补全。trace 持久化改为少量批量写入,单次 cognify 运行即可算出 pass 窗口,数据库写入和重复流水线运行都少了,高流量会话的入库延迟明显下降;会话与数据集的绑定也更可靠,技能运行能正确路由到绑定数据集,修掉若干因绑定过期或错配引发的报错。
新功能:函数、类与模块的 docstring 现在作为一等文档存入代码图并与代码符号关联,搜索和代码智能体因此能给出带上下文的答案,并在返回代码引用时附上内联文档;新增 Memify API 及 CLI 接线,可直接把文件、消息等内容转成记忆条目;在供应商支持时改用 base64 编码请求向量,改善互操作性、减少二进制传输故障。安全与连接器方面加入了每日 CVE Slack 告警、Google Calendar 连接器加固以及 OAuth 与令牌共享的改进。
破坏性变更:若自定义 graph_model 生成步骤,需将 datamodel-code-generator 升到 0.56.0 及以上,否则会报生成错误。
v1.6.5.dev0 — Distillation, Embeddings & Connector Hardening2026-10-16
v1.6.5.dev0 主打记忆蒸馏、向量嵌入与连接器加固。蒸馏流程更精准:只读取所需的 QA 轮次而非整段对话,按批次封存中间结果,并默认对新增记忆执行蒸馏,摘要更稳定;结构化记忆条目现在会正确挂到当前会话的数据集下,避免落错库。嵌入侧支持按需请求 base64 编码向量、计量 token 用量、默认缩小批量以换稳定性,本地模型缓存统一存放于 ~/.cognee/models。Gmail、Google Calendar 连接器的配额与权限处理更稳健,litellm、Mistral 调用与 Azure 令牌逻辑也做了整合。
另有两项实用功能:仓库摄入时可指定克隆分支,以及每日 CVE 告警自动推送到 Slack。
升级前注意三处破坏性变更:旧版 triplet 嵌入接口已弃用并部分移除,需迁移到新的 embedding/memify 管线;自定义图模型要求 datamodel-code-generator ≥ 0.56.0;FASTEMBEDCACHEPATH 已从 .env.template 移除,自定义过该路径的用户请手动迁移缓存目录。
v1.6.4 — Vector storage & DB reliability fixes2026-10-13
v1.6.4 是个纯修复版本,发布于 2026-10-13,主要收口向量检索与数据库适配层的几处可靠性问题。LanceDB 和 pgvector 适配器搜索时不再顺带取出原始向量列,除非显式要求——大块嵌入数据既占内存又拖慢响应,此前却是每次查询的固定开销,现在网络与内存占用都降下来了。Turso 打开失败时统一抛出 SQLAlchemy 的 OperationalError,调用方用同一套异常处理即可覆盖各数据库后端,不必再单独捕获底层客户端异常。需要注意一个行为变化:LanceDB 存储移除了后台自动清理任务,以免搜索和测试期间冒出意外的并发写入;若你原本依赖自动清理回收磁盘,得改用 LanceDB 工具或既有运维流程手动执行 compaction。此外补了若干测试、修了测试框架的子进程与会话问题,缓解 CI 偶发抖动。锁文件同步至 cognee 1.6.3,Cognee-MCP 升到 0.5.7,lancedb、lance-namespace、pylance 等依赖同步更新。本版无新增功能,也无安全相关改动。
v1.6.4.dev0 — Vector search reliability & Turso error handling2026-10-13
v1.6.4.dev0:修复向量搜索取回原始向量列的问题,Turso 打开文件失败时统一抛出 OperationalError。
这版主要解决两个痛点。一是 pgvector 和 Lancedb 适配器在相似度搜索时误选了原始向量列,导致响应体过大甚至序列化报错;现在搜索只返回元数据和近邻分数,网络开销和内存占用都降下来了。二是 Turso 适配器把底层的 pyturso IoError 包装成 SQLAlchemy 的 OperationalError,上层应用和 ORM 不必再处理各种意外异常类型。
配套更新了 lancedb(0.39.0)、pylance(12.0.0)等锁文件依赖,cognee-mcp 升到 0.5.7,并补充了向量距离语义与 Turso 行为的测试。