Agent Context Management: Open-Source Solutions Compared

原创

克劳德2048

发布于 2026-10-22 16:59:07

1630

Agent context management has become its own category in 2026. It's no longer just "memory" — it's about what context enters the agent's window, how it's compressed, who controls it, and whether it survives across sessions and frameworks.

Here's a comparison of the main open-source options.

The Landscape

Solution	Memory Types	Framework Coupling	Team Sharing	Context Compression	License
Mem0	Conversation summary	Low (SDK)	Weak	Vector retrieval	Apache 2.0
Letta (MemGPT)	Conversation + summary	Low (SDK)	Weak	Pagination	Apache 2.0
Zep / Graphiti	Conversation + knowledge graph	Low (SDK)	Weak	Graph retrieval	Apache 2.0
LangMem	Conversation summary	High (LangChain)	None	Framework-level	MIT
TencentDB Agent Memory	Chat Memory + Skill + Wiki + CodeGraph	Low (Proxy, zero-code)	Team-level + ACL	L0-L3 layered + BM25/vector/RRF + budget caps	MIT

What Sets TencentDB Agent Memory Apart

1. Four asset types, not just conversation summaries

Most solutions store one thing: summarized conversations. TencentDB Agent Memory recognizes that agents reuse different types of information:

Chat Memory — who the user is, what they've decided, what they prefer
Skill — proven workflows with versions, triggers, and validation rules
Wiki — documents structured into pages with a link graph
CodeGraph — code symbols, call relationships, and impact paths

2. Framework-agnostic via Proxy

One Proxy, unchanged protocol. Point any agent's base URL to the Proxy — done. No plugin, hook, or MCP server. Currently supports: DeepSeek Harness, Claude Code, Codex, CodeBuddy, WorkBuddy, Hermes, OpenClaw.

Memory assets move freely across frameworks. Switch from Claude Code to Codex? Your memory follows.

3. Budget-controlled retrieval

Retrieval results are constrained by entry count, character budget, and timeout. Memory enriches context without overflowing it — a critical feature for token cost management.

The four-layer architecture stores at different granularities:

Layer	Stores	Purpose
L0 Conversation	Raw conversations	Source verification
L1 Atom	Extracted facts, preferences, constraints	Precise recall
L2 Scenario	Project/scenario knowledge blocks	Fast context recovery
L3 Persona	Long-term user profile	Quick understanding

Retrieval: BM25 + vector search + RRF, with hard constraints.

4. Team memory with access control

Three visibility levels: private (owner-only, even admins can't read), team (members read), restricted (User/Role/Agent ACL). Skills go from private → reviewed → team-shared → assigned to specific agents.

New Chat Memory and Skills are private by default. Sharing is an explicit action, not a default leak.

5. Cold-start: import your existing assets

Codebases → CodeGraph auto-indexes symbols, files, call relationships
Documents → Wiki auto-generates structured pages with link graph
Past conversations → Skills and Chat Memory auto-extracted

Stop retraining every agent. Give it the save file.

Memory Assets vs Chat History vs Standard RAG

	Chat History	Standard RAG	TencentDB Agent Memory
Cross-session user understanding	△	△	✅ Chat Memory
Distilled executable experience	—	—	✅ Skill
Document structure & relationships	—	△ Chunk retrieval	✅ Wiki + Link Graph
Code call graphs & impact scope	—	△ Text match	✅ CodeGraph
Ownership / Version / Status	—	—	✅
Team sharing & Agent loadout	—	—	✅
Private / Team / ACL	—	△	✅

Benchmark

PersonaMem: 48% → 76% (+59%) with memory enabled.

Quick Start

git clone https://github.com/TencentCloud/TencentDB-Agent-Memory.git
cd TencentDB-Agent-Memory/deploy/global-images
cp .env.example .env
$EDITOR .env       # Fill in two sets of LLM parameters
./start-all.sh     # Launch everything with one command

Panel: http://localhost:8125

Repo: https://github.com/TencentCloud/TencentDB-Agent-Memory

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

chat