How ContextBucket Gives Agents Unlimited Memory and a Unified Workspace

Background and Challenges

When agents move from demo chats to production systems—writing code, performing analysis, handling tickets—they become long‑running processes that need persistent context. Three major gaps emerge:

Memory gap: Vector stores or local SQLite lose data when devices change; pure vector retrieval is noisy and imprecise.

Workspace gap: Code, configs, docs, and artifacts stored locally disappear on a new machine and cannot be shared across teams.

Governance gap: Multiple agents share data without native isolation, leading to permission, audit, and quota problems.

Product Capability: One Base, Two Abilities

ContextBucket, a Volcano Engine managed service, provides file storage, memory management, hybrid retrieval, multi‑tenant isolation, and serverless elasticity within a single service layer. Agents integrate via a ContextBucket plugin, receiving all capabilities without separate vector‑store or object‑storage connections. Each agent or user gets a ContextSet (logical isolation unit) where memory and workspace share credentials and endpoints.

Memory: Remember Accurately, Find Quickly, Carry Anywhere

Accurate recall ("记得准"): ContextBucket extracts only key facts—decisions, technical conclusions, user preferences—filtering out process noise and handling relative time expressions to avoid ambiguity.

Example: After a two‑day OpenClaw conversation about microservice design, the system stores only the final decisions (domain‑split architecture, gRPC choice, error‑code conventions) instead of the entire discussion.

Precise retrieval ("找得到"): Uses a dual‑path search (vector + BM25) with reranking. The first session triggers a broad recall to avoid missing context from vague queries.

Example query: "Why was gRPC chosen for the core link?" Pure vector search returns many noisy fragments, but hybrid search surfaces the single decisive statement.

Portability ("带得走"): Memory is stored server‑side. Switching machines or environments requires only the same user_id; the full memory becomes instantly available. In Locomo evaluations, this reduced LLM output tokens by 80% and billing tokens by 43.2%.

Workspace: Persistent Files, Continuous Workflow

Files (code, configs, docs, artifacts) are persisted remotely. With FUSE mounting, the workspace appears as a local file system, allowing existing tools to operate unchanged while data lives in ContextBucket.

Example: A data‑processing script saved in the workspace remains after a computer reboot and is instantly available on another machine; teammates can reuse the script without manual transfer.

Workspace retrieval combines three parallel paths:

Vector search for semantically similar code or documentation.

BM25 keyword search for exact function names, config keys, error codes.

Path/metadata matching for directory, filename, and modification‑time filters.

Rerank merges results, and only the most relevant snippets are injected according to the token budget, keeping large repositories stable.

Example: Querying "Which config sets the traffic ratio for gray‑release?" returns the exact line instead of dozens of similar files.

Workspace and memory can be returned together in a single retrieval request, simplifying agent logic.

Integration Model

Installation requires two steps: one‑click plugin install and verification. The following command installs the plugin (replace placeholders with real credentials):

curl -fsSL https://context-bucket-cn-beijing.tos-cn-beijing.volces.com/context-bucket-bundle-latest.tar.gz | tar xz -C /tmp && bash /tmp/stage/install.sh \
   --backend context \
   --endpoint tos-control-cn-beijing.volces.com \
   --access-key-id '<YOUR_VOLC_AK>' \
   --access-key-secret '<YOUR_VOLC_SK>' \
   --region cn-beijing \
   --account-id '<YOUR_ACCOUNT_ID>' \
   --context-bucket-name 'context-bucket-poc' \
   --context-set-name 'csn-poc' \
   --secure false \
   --force

After installation, run openclaw plugins list to verify the plugin status.

Performance Validation: Locomo Long‑Conversation Benchmark

Locomo is a standard long‑dialog benchmark measuring memory across sessions. The baseline OpenClaw version and the ContextBucket‑enabled version use the same model and question set; the only difference is the storage backend.

Results: answer accuracy rose from 16.45% to 64.14% (an increase of 48 percentage points). The core reason is the extract‑→‑retrieve‑→‑filter pipeline that injects only relevant snippets, cutting LLM output tokens by 80% and billing tokens by 43.2%.

Applicable Scenarios

R&D agents: Long‑term code, design docs, CI results, and troubleshooting knowledge need cross‑device continuation and team sharing.

Office / workflow agents: Meeting minutes, weekly reports, approvals, and historical decisions require stable long‑term context.

Personal assistants: User preferences, historical tasks, and local files must persist across phone, PC, and car.

Enterprise Copilot platforms: Multi‑agent, multi‑team, multi‑tenant environments need native isolation, quota management, and auditability.

Summary and Outlook

ContextBucket consolidates the three context gaps—memory loss, workspace inconsistency, and governance chaos—into a single managed base. Empirical results on OpenClaw show:

Memory gap solved: vector + BM25 + rerank retrieval, server‑side persistence, accuracy 16.45% → 64.14%.

Workspace gap solved: FUSE mount, semantic + keyword + metadata multi‑path retrieval, seamless cross‑agent workflow.

Governance gap solved: ContextSet provides ten‑million‑scale native multi‑tenant isolation, unified credentials, clear audit trail, serverless zero‑ops deployment.

Future directions include smarter retrieval with graph and temporal awareness and an open ecosystem that expands plugin support (LangChain, Hermes, etc.) and exposes ContextSet APIs for third‑party tools.

In short, ContextBucket enables agents to remember, find, and carry their memory and workspace on a unified, server‑side foundation.