Changelog

All notable changes to the building-with-zep plugin. This file is the only release feed the plugin system offers — Claude Code, Codex, and Cursor have no changelog plumbing, so users who want to know what a version changed read this.

The version here is the release identity shared by the portable and vendor manifests; Claude Code also uses it for update detection. Bump it with python3 scripts/plugin_manifests.py set <version> from the repo root; see Releasing.

0.10.0 — Unreleased

Added

• Add the migrating-zep-v3-to-v4 skill at skills/migrating-zep-v3-to-v4/SKILL.md.
• Link to the migration skill from the building-with-zep skill.

Changed

• The building-with-zep skill now targets the Zep V4 API and SDKs.

0.9.1 — 2026-10-14

Added

• Add an "Intake" step to the tuning-zep-ingestion procedure, before Step 0. The agent asks how the graph will be used, which questions the graph must answer, and whether the customer has evaluation question-and-answer pairs. The agent also asks the user to upload a sample of the data. The step includes an example message to the user and an example question-and-answer CSV file. Step 0 refers to the Intake step for the gold question set.

0.9.0 — 2026-10-08

Added

• Add the tuning-zep-ingestion skill again, at skills/tuning-zep-ingestion/SKILL.md. The building-with-zep skill references are the same as in version 0.7.0.

Changed

• Update the tuning-zep-ingestion procedure from version 0.7.0. A new "API surface" table names the operations of each step without SDK method names. The procedure gives the imported node the name that the prose uses, and it refers to edge endpoints by node UUID. The lead sentence keeps the subject identifier and drops the identifiers of related records. Step 4 records the debug-mode result in the manifest and checks the batch debug logs through the API. Step 5 measures three kinds of field repeats, and Step 6 compares the trace totals with the graph totals.
• State the rule for the split in the tuning-zep-ingestion procedure. Structured data goes into the deterministic import, through one field mapping for each structured schema. Unstructured data becomes narrative episodes. The agent must not use an LLM to make structured records from prose for the import.

0.8.0 — 2026-10-07

Removed

• Remove the tuning-zep-ingestion skill that version 0.7.0 added. The plugin content is the same as in version 0.6.0. The building-with-zep skill is again the only skill in the plugin.

0.7.0 — 2026-10-07

Added

• Add the tuning-zep-ingestion skill at skills/tuning-zep-ingestion/SKILL.md. It gives the procedure to turn a dataset into a domain-specific Context Graph, from data splitting through versioned ingestion and graph measurement.
• Tell the agent in the building-with-zep skill to use the tuning-zep-ingestion skill for ingestion code and tuning. The ingestion skill tells the agent to activate the building-with-zep skill first and use zep-docs if that skill is not available.

0.6.0 — 2026-10-04

Changed

• Teach two retrieval models in the skill. Context Block grounding uses one low-latency call for each turn. An agent with Zep tools learns the graph, uses domain knowledge, makes a retrieval plan, and runs several targeted tool calls for complex questions and tasks.
• Add an invariant: an agent with one search tool and no graph orientation or domain knowledge gets related facts of low value.
• State that application-authored domain knowledge can go in the system prompt, and that retrieved graph content never goes there.
• Limit the "favor recall over precision" rule to Context Block grounding.
• Add agent evaluation guidance to "Evaluating Zep".
• Index the new "Build an Agent with Zep" and "Build Tools for an Agent" pages. Replace the link to the removed "Retrieval philosophy" page.

0.5.2 — 2026-10-03

Changed

• Rename the Memory MCP server to the Context MCP server and update its documentation link.

0.5.1 — 2026-10-01

Fixed

• Fix the building-with-zep skill frontmatter. The description value was an unquoted YAML scalar that contained a colon, so strict YAML parsers rejected the skill. The value is now a folded block scalar and is shorter than the 1024-character limit.

0.5.0 — 2026-09-24

Changed

• Position Zep as the unified context layer for enterprise data in the building-with-zep skill, and describe agent memory as one of three solutions: enterprise context graphs, agent memory, and customer and account context. Add a "Choose the solution" section that maps each solution to its graph scope and retrieval path.
• Use "shared Context Graph" for the base graph type, and note that the SDK and API still call it a standalone graph.
• Extend the implementation lifecycle to Ingest → Shape the Graph → Retrieve → Govern, with a new section on access policies (RBAC, ABAC, source traceability, Audit Logs, and API Logs) and guidance that retrieved context is evidence, not instruction.
• Reorganize the documentation index to follow the new docs layout and add the pages that were missing (context type pages, fact-retrieval cookbooks, Memory MCP server, rate limits, HIPAA, BYOM, Zep vs Graphiti, zepctl, FAQ).
• Replace end-user trigger phrases with developer implementation phrases, starting with "implement Zep for my agent".

0.4.3 — 2026-08-26

Changed

• Make assets/logo.png exactly square (704×704) so it meets Cursor Marketplace’s 1:1 logo requirement.
• Document multi-graph backfill guidance in the building-with-zep skill: submit all episodes to all graphs without waiting for processing between graphs or sources, and poll only once on the last episode when retrievability matters (applies to thread.add_messages, graph.add, batch, and zep-ingest alike).

0.4.2 — 2026-08-18

Added

• Document plugin-specific terms (Apache-2.0) and privacy (docs MCP only; not Zep Memory) in legal/TERMS.md and legal/PRIVACY.md, linked from the README. Product privacy/terms apply only if you later use Zep’s APIs.
• Point plugin support to support@getzep.com in the README.

Changed

• Drop the Cursor vendor plugin wrapper (.cursor-plugin/plugin.json and .cursor-plugin/mcp.json). Cursor loads the Agent Plugins package. The Cursor marketplace catalog stays and still points at this repo root.
• Rewrite the plugin Legal section as formal Terms of Service and Privacy Policy copy (Apache-2.0 still governs the plugin files; product terms still apply only if you later use the Zep Service).
• Move those terms and the privacy policy into legal/TERMS.md and legal/PRIVACY.md. Remove the statement that Zep does not use Plugin information to train models.

0.4.1 — 2026-08-17

Added

• Commit the Zep icon at assets/logo.png and reference it from the Cursor plugin and marketplace manifests.
• Document plugin usage and configuration in README.md.

0.4.0 — 2026-08-12

Changed

• Host Claude Code and Codex marketplaces in this repository with same-repo plugin sources. Install with claude plugin marketplace add getzep/building-with-zep-plugin then claude plugin install building-with-zep@building-with-zep.
• Update the building-with-zep skill ingestion guidance: document zep-ingest and when to use it vs. thread.add_messages / graph.add, point agents at Create an ingestion pipeline and Prepare data for ingestion, and remove the retired JSON best-practices page.

0.3.0 — 2026-08-07

Added

• Add a root plugin.json and schema-qualified mcp.json conforming to Agent Plugins 1.0.0, so compatible clients can load the shared skill and Zep Docs MCP server without vendor-specific adaptation.
• Add dependency-free conformance checks to CI.

Changed

• Retain and validate the Claude, OpenAI, and Cursor package manifests alongside the portable package. Claude and OpenAI installs continue to use their vendor wrappers when the client does not load Agent Plugins directly.

0.2.0 — 2026-07-30

Added

• Cursor as a third ecosystem (#581). The plugin directory now carries a .cursor-plugin/plugin.json manifest and a Cursor-shaped mcp.json, and the repo root carries a .cursor-plugin/marketplace.json. Install with /add-plugin building-with-zep@https://github.com/getzep/zep in a Cursor 2.5+ Agent chat. All three runtimes load the same skills/building-with-zep/ tree.

Changed

• The version now lives in exactly one place per ecosystem, and CI fails on drift. version was removed from the building-with-zep entries in both .claude-plugin/marketplace.json and .agents/plugins/marketplace.json: each ecosystem resolves the release from its own plugin.json, so marketplace copies are dead weight that can mask a bump. The Claude marketplace manifest's own metadata.version was removed for the same reason — nothing consumes it, and it reads as the plugin's version.
• CI now requires a strictly newer semantic version whenever loaded plugin content changes. The release helper updates all three manifests together and refuses invalid versions or downgrades.
• All three manifests now link to the canonical plugin documentation, and the Codex manifest includes the interface metadata required by current plugin ingestion.

0.1.0 — 2026-06-15

Initial release (#520): the building-with-zep skill plus the zep-docs MCP server (https://docs-mcp.getzep.com/mcp, remote HTTP, no API key).

Two later changes shipped under this same 0.1.0 string, so anyone who installed before them never received them and had to reinstall:

• Codex as a second ecosystem, sharing one skill and MCP config (#566).
• Restored skill guidance and the "what goes in the skill vs. the docs" placement rule (#567).

That is the failure mode 0.2.0's tooling exists to prevent: Claude Code uses the explicit version for update detection, so pushing plugin content without increasing it leaves existing Claude Code installs on the old cached release.