LOCAL ALPHA OPENREST APIMCP STDIOPOSTGRES 16DASHBOARDPRIVATE DEPLOY

Lore Context.

AI Agent 记忆、评估和治理的控制平面。

在记忆变成生产风险之前，看清每个 agent 记住了什么、使用了什么、应该忘记什么。

运行本地 Alpha → View on GitHub ↗

$ pnpm quickstart -- --dry-run --activation-report

Golden clients3

MCP tools11

Demo eval size4

Latencypending

lore › context.ledger local · demo fixture

Demo memories4public fixture

Golden clients3Claude · Cursor · Qwen

Trace rows2retrieved and used

P95 latencypendingnot claimed

sourcememoryevidenceused_in_responsestale_scoresensitivityreview_status

claude-code#a4f1api-contractyes lowapproved

cursor#b2c8pricing-noteyes lowapproved

qwen-code#c9d3deploy-fixyes medflagged

hermes#d1a7private-keyyes highredact

fastgpt#f3c1old-schemano medreview

context.compose()
demo eval
policy pass
trace linked
k 5

gov · gate

demo-private · audit_loglocal alpha · public fixture

Multiple coding agents

Share governed context across Claude Code, Cursor, Qwen Code, and local agents.

Audit agent memory

Open a trace and see which rows were retrieved, used, ignored, stale, or risky.

Migrate memory later

Use MIF-style export fields for provenance, supersedes, and contradicts relationships.

01 / PROBLEMThe memory gap

Agent 会记忆，团队需要证据。

长期记忆正在进入工作流，但多数团队仍无法解释一次回答背后的审计路径。

过期记忆

过去的 API 约定可能混入今天的输出。

出处不明

事实看似有用，却不知道来自哪个会话。

危险写入

幻觉、密钥和敏感上下文可能被写入共享记忆。

后端锁定

不同向量库用不同 schema、格式和检索路径锁住数据。

02 / SYSTEMPipeline

Agent 和记忆之间的治理路径。

Lore 评估检索、追踪回答路径，并把高风险写入送入治理审查。

System diagram · context plane v0.6read path · policy gate · persist + audit

01MCP clients

Lore 暴露 MCP stdio 和 REST 接口，让查询、写入、审查都经过同一治理层。

02context.query

带出处、新鲜度和策略状态的排序记忆检索。

03composer

Lore 评估检索、追踪回答路径，并把高风险写入送入治理审查。

04eval / trace / governance

调整检索器、重排器和新鲜度阈值时，观察 recall、precision、stale-hit 和 latency 的变化。

05Postgres + adapter

导出记忆、出处、嵌入和策略状态。

03 / FEATURESProduct surfaces

六个产品界面，一条审计链路。

没有玄学记忆层；每个能力都可检查、可脚本化、可关闭。

F · 01context.query

Context Query

带出处、新鲜度和策略状态的排序记忆检索。

stdio · httpk=1..32ttl-aware

F · 02eval.run

Eval Playground

用种子数据重放并比较检索设置。

recall@5precision@5MRR

F · 03observability

Memory Observability

追踪每次检索、写入、删除和策略决策。

span treeper-agentp50/p95/p99

F · 04governance

Governance Review

审批、修订、拒绝或遗忘候选记忆。

policy.loreapprove/rejectdiff view

F · 05export.mif

Memory Portability

导出记忆、出处、嵌入和策略状态。

JSON Lines.lore.tarreplay()

F · 06deploy.local

Private Deployment

用 Docker Compose 和 Postgres 在本地或 VPC 运行。

docker composeair-gappedBYO Postgres

04 / ALPHAShipped this week

v0.6 alpha 已包含的能力。

先真实可用，再持续打磨。以下内容都可以通过 Docker Compose 在本地运行。

✓Graph-Augmented Hybrid Memory Engine betaPostgres FTS + pgvector hybrid recall, v12 episode/item/chunk schema, Postgres temporal graph, lifecycle propagation, v12 eval gate / production gate1.2.0-beta

✓Cloud Memory Engine betaGoogle sign-in, automatic capture, SQS worker fan-out, RDS truth store, Qdrant hybrid recall, Memory Inbox exceptions, and Evidence Ledger1.1.0-beta

✓REST API + OpenAPI/v1/context, /v1/memory, /v1/eval — documented at /openapi.jsonv0.6.0

✓MCP stdio SDK transportDrop-in for any MCP-compatible clientmcp 0.6

✓Evidence LedgerTrace-centric proof of retrieved, used, ignored, and risky memoryledger

✓Activation reportDry-run and real first-value proof with redacted local evidenceactivation

✓AI-readable docs/llms.txt, /llms-full.txt, canonical metadata, and public-source boundariesllms

✓Public-safe evalShare metrics without raw memory content or dataset messageseval

✓Distribution packMCP registry, agent plugin, marketplace, and launch content draftsdocs

✓Design partner intakeIssue template, intake rubric, and activation scorecardintake

◐Public SaaSHosted multi-tenant sync, billing, and remote MCP HTTP stay out of v0.6roadmap

Build manifest

1.2.0-beta · 04.05.2026

graph beta: v1.2 graph-augmented hybrid memory engine
public release: v1.2.0-beta formal pre-release
runtime: node 22+ · RDS 16 · Qdrant
smoke: passing · production smoke
bundle: static website · Docker images
memory floor: RDS truth · Qdrant index
telemetry: privacy-safe
license: Apache 2.0

# first run

git clone github.com/Lore-Context/lore-context
pnpm install && pnpm quickstart -- --dry-run --activation-report
pnpm openapi:check # release gate

05 / EVALMethodology

在你自己的数据上证明记忆质量。

调整检索器、重排器和新鲜度阈值时，观察 recall、precision、stale-hit 和 latency 的变化。

prod/smoke-2026-05-04OpenAPI · Hosted MCP · OAuth start · Dashboard · LLM files · no benchmark claimPUBLIC · v1.2.0-beta · memory engine

Recall@51.0004-item smoke dataset vs baseline

Precision@50.200expected with k=5 vs baseline

MRR0.875one query ranked #2 vs baseline

Stale-hit0.000no stale demo rows vs baseline

p95 latencypendingnot measured via MCP vs baseline

What the smoke run proves

pipeline retrieves gold memorypass

Evidence Ledger trace is inspectablepass

stale rows are reported as zeropass

larger dataset still requiredtodo

latency must be measured via direct APItodo

Publication guardrails

4-item datasetsmall

no LOCOMO claimtrue

no p95 claimtrue

raw data archivedyes

methodology firstyes

Use this as methodology, not a benchmark win.Read report →

06 / INTEGRATIONSDrop-in MCP

接入你已经在用的 agent 协议。

Lore 暴露 MCP stdio 和 REST 接口，让查询、写入、审查都经过同一治理层。

CLClaude Codemcp.stdio · alpha

CUCursormcp.stdio · alpha

QWQwen Codemcp.stdio · alpha

OPOpenClawmcp.stdio · alpha

HEHermesmcp.stdio · alpha

DIDifyrest.v1 · alpha

FAFastGPTrest.v1 · planned

RORoo Coderest.v1 · planned

OPOpenWebUIrest.v1 · planned

LOLocal agentsrest.v1 · planned

07 / STARTStart

从本地 Alpha 开始，在扩展之前证明记忆质量。

四条命令、种子演示数据集，以及验证 dashboard 正常渲染的 Playwright smoke。

运行本地 Alpha →查看架构

$ git clone https://github.com/Lore-Context/lore-context
$ cd lore-context
$ pnpm install
$ pnpm quickstart -- --dry-run --activation-report
$ pnpm openapi:check
$ pnpm smoke:dashboardverified · OpenAPI · evidence · smoke passing